Machine learning-based adhesive coating uniformity optimization control method and system

Abstract

The invention relates to the technical field of adhesive coating automation control, and discloses an adhesive coating uniformity optimization control method and system based on machine learning. The method comprises the following steps: acquiring real-time process parameters and online images in a coating process and acquiring quality monitoring data acquired by an infrared sensor; and inputting the data into a pre-trained quality prediction model to obtain a coating uniformity prediction value and a potential defect area identifier. A first control adjustment amount used for adjusting the coating speed and rolling is generated according to the difference value between the uniformity predicted value and the target; and generating a second control adjustment amount for adjusting the tension of the base material and the viscosity of the adhesive according to the potential defect area identifier. And fusing the two to form a comprehensive control instruction and issuing for execution. According to the method, advanced prediction of the coating quality and accurate positioning of the local defects are realized, the global uniformity and the local defects are synchronously optimized through the sub-target cooperative control strategy, and the stability and consistency of the coating quality are improved.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for adhesive coating, specifically to a machine learning-based method and system for optimizing and controlling the uniformity of adhesive coating. Background Technology

[0002] In the adhesive coating production process, coating uniformity is a core indicator determining product quality. Current technologies primarily rely on single-type sensors for offline or online inspection of the coated adhesive layer, such as using beta-ray or infrared sensors to measure the thickness or degree of cure at fixed points. Operators then manually adjust individual process parameters like coating speed and roller pressure based on this limited point data or defects discovered during subsequent quality inspections, relying on experience. This control method suffers from significant lag, failing to intervene in the early stages of defect formation, and lacks a systematic approach to adjustment.

[0003] Existing control methods struggle to perceive the coating status in real time and comprehensively. Minor thickness unevenness or incomplete curing in localized areas is easily missed in point-based detection, and subsequent adjustments are often general operations applied to the entire equipment, failing to precisely target defective areas. This leads to significant adhesive waste during production, poor product consistency, and bottlenecks in improving the yield of high-quality products. Therefore, the production floor requires a technical solution capable of predicting uniformity changes in advance and accurately locating potential defective areas. Simultaneously, transforming this predictive and location information into differentiated and coordinated control commands for different actuators to achieve simultaneous optimization of global uniformity and localized defects presents another key technical challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based method and system for optimizing and controlling the uniformity of adhesive coating, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a machine learning-based method for optimizing and controlling the uniformity of adhesive coating, the method comprising: Collect a set of real-time process parameters during the adhesive coating process, including coating speed, application roller pressure, substrate tension, and adhesive viscosity data. Receive quality monitoring data of the coating area, the quality monitoring data including the adhesive layer thickness distribution map captured by the online image sensor and the adhesive layer curing degree distribution map detected by the infrared sensor; The real-time process parameter set and the quality monitoring data are input into the pre-trained quality prediction model. Through the calculation of the quality prediction model, the predicted value of coating uniformity and the identification of potential defect areas are output. Based on the difference between the predicted coating uniformity value and the preset target uniformity, a first control adjustment amount is generated, which is used to adjust the coating speed and the pressure of the application roller. Based on the identification of the potential defect area, a second control adjustment amount is generated, which is used to adjust the substrate tension and the viscosity of the adhesive; A comprehensive control command is generated based on the first control adjustment amount and the second control adjustment amount, and the comprehensive control command is sent to the coating equipment for execution to control the coating process of the next production cycle.

[0006] Preferably, the step of inputting the real-time process parameter set and the quality monitoring data into a pre-trained quality prediction model, and outputting a predicted coating uniformity value and potential defect area identifiers through the calculation of the quality prediction model, includes: The thickness distribution map of the adhesive layer is subjected to feature extraction processing to obtain a thickness distribution feature vector, which includes the average thickness, thickness range, thickness variance and thickness change trend slope. The curing degree distribution map of the adhesive layer is processed by feature extraction to obtain a curing degree distribution feature vector, which includes the average curing degree, curing degree uniformity coefficient and coordinates of curing degree anomaly points; The thickness distribution feature vector, the curing degree distribution feature vector, and the real-time process parameter set are normalized and fused to generate a comprehensive feature matrix containing multi-dimensional time series. The comprehensive feature matrix is ​​input into the deep neural network of the quality prediction model, the deep neural network comprising a long short-term memory network layer and an attention mechanism layer; The deep neural network calculates and generates a predicted value for the coating uniformity of the entire coating area. It also outputs the coordinates of areas with excessively thin thickness, uneven curing, or risk of bubble formation on the adhesive layer thickness distribution map, which serve as identifiers of potential defect areas.

[0007] Preferably, the step of performing feature extraction processing on the adhesive layer thickness distribution map to obtain a thickness distribution feature vector includes: The thickness distribution map of the adhesive layer is divided into multiple uniform grid units; Calculate the average thickness of the adhesive layer for each grid cell, and record the maximum and minimum thickness values. The thickness range is calculated by the difference between the maximum and minimum thickness values. The average thickness distribution of all the grid cells is statistically analyzed, and the standard deviation of the average thickness is calculated as the thickness variance. Along the coating direction, the average thickness sequence of multiple continuously arranged grid cells is linearly fitted, and the slope of the fitted line is used as the slope of the thickness change trend.

[0008] Preferably, generating a first control adjustment amount based on the difference between the predicted coating uniformity value and the preset target uniformity includes: Calculate the absolute difference between the predicted coating uniformity value and the preset target uniformity; The absolute difference is input into the control strategy model, which includes a nonlinear relationship function between coating speed and application roller pressure. Calculated by the control strategy model, when the absolute difference exceeds the first threshold, a first speed adjustment amount to reduce the coating speed and a first pressure adjustment amount to increase the pressure of the application roller are generated. When the absolute difference is lower than the second threshold, a second speed adjustment amount to increase the coating speed and a second pressure adjustment amount to decrease the pressure of the application roller are generated. The first control adjustment amount is formed by combining the first speed adjustment amount with the first pressure adjustment amount, or by combining the second speed adjustment amount with the second pressure adjustment amount.

[0009] Preferably, the step of inputting the absolute difference into the control strategy model, wherein the control strategy model includes a nonlinear relationship function between the coating speed and the application roller pressure, including: A dynamic relationship model between coating speed and application roller pressure is established. The dynamic relationship model defines the mapping relationship between the change in coating speed and the change in application roller pressure when maintaining a constant coating amount per unit area under a specific adhesive viscosity. In the dynamic relationship model, the absolute difference is introduced as a feedback factor, and the proportional coefficient of the mapping relationship is dynamically adjusted by the sign and magnitude of the absolute difference. The dynamic relationship model is trained using offline training data to obtain a control strategy model containing a corrected proportional coefficient, which is used to accurately calculate the adjustment amount of speed and pressure based on the absolute difference input in real time.

[0010] Preferably, generating the second control adjustment amount based on the potential defect area identifier includes: The potential defect area identifier is analyzed to identify the location distribution of the potential defect area and the defect type in the coating width direction; When the defect type is too thin, the substrate at the corresponding position of the potential defect area needs to be supplemented with tension, and the substrate tension increase adjustment amount is generated. When the defect type is uneven curing or bubbles, the database relating adhesive viscosity and curing properties is queried to calculate the target adhesive viscosity value required to improve the curing state of the potential defect area and generate a viscosity adjustment instruction. The second control adjustment amount is formed by combining the substrate tension increase adjustment amount with the viscosity adjustment command.

[0011] Preferably, the step of calculating the substrate tension required at the location corresponding to the potential defect area to generate an increased substrate tension adjustment includes: Determine the center coordinates and width of the potential defect area along the coating width direction; Based on the center coordinates and the area width, a local tension compensation coefficient is calculated, which is proportional to the area width. Obtain the current overall tension value of the substrate, and multiply the overall tension value of the substrate by the local tension compensation coefficient to obtain the additional tension required for the substrate; The additional tension required for the substrate is converted into a torque control increment on a specific tension roller, resulting in an adjustment amount for increasing the substrate tension.

[0012] Preferably, the step of generating a comprehensive control command based on the first control adjustment amount and the second control adjustment amount includes: Determine whether there is an execution conflict between the first control adjustment amount and the second control adjustment amount; When there is no execution conflict, the first control adjustment amount and the second control adjustment amount are directly merged to generate the comprehensive control instruction; When execution conflicts exist, the parameters in the first control adjustment amount or the second control adjustment amount are adjusted according to the preset conflict resolution rules. The conflict resolution rules prioritize ensuring the uniformity of coating thickness and then consider the uniformity of curing. The adjusted control values ​​are combined to generate the final integrated control command.

[0013] Preferably, the steps for constructing the pre-trained quality prediction model include: Collect multiple sets of process parameter data and corresponding quality test results data during the historical adhesive coating process. The process parameter data includes coating speed, application roller pressure, substrate tension and adhesive viscosity. The quality test results data includes adhesive layer thickness distribution measurement value and adhesive layer curing degree distribution measurement value. The historical process parameter data and quality inspection result data are cleaned and labeled to form a training sample set; A deep neural network model is constructed as the basic structure of the quality prediction model. The deep neural network model includes an input layer, multiple hidden layers, and an output layer. The deep neural network model is trained using the training sample set, and the model parameters are optimized using the backpropagation algorithm until the error between the model's predicted output and the quality detection result data is lower than a preset threshold. The trained deep neural network model is saved as the pre-trained quality prediction model, which is used to receive the real-time process parameter set and the quality monitoring data in real time and output the prediction results.

[0014] Preferably, the present invention also includes a machine learning-based adhesive coating uniformity optimization control system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the machine learning-based adhesive coating uniformity optimization control method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The pre-trained quality prediction model simultaneously receives and processes a set of numerical real-time process parameters and image-based online quality monitoring data. By fusing and analyzing coating speed, pressure, tension, viscosity, adhesive layer thickness distribution maps, and curing degree distribution maps, the model calculates predicted coating uniformity values ​​and identifies potential areas where coating may be too thin, too thick, or insufficiently cured. This enables proactive judgment and spatial location of coating quality, changing the passive approach of traditional methods that rely on post-event sampling and single-point measurements. It allows the control system to obtain early warnings and defect location information before uniformity actually deteriorates, creating conditions for precise control.

[0016] Based on the model's output of predicted coating uniformity and identified potential defect areas, a first control adjustment value and a second control adjustment value are generated. The first control adjustment value addresses the predicted overall uniformity deviation by primarily adjusting the coating speed and application roller pressure to correct macroscopic deviations in adhesive application. The second control adjustment value targets specific defect areas by selectively adjusting the substrate tension and adhesive viscosity, which affect the leveling and adhesion of the adhesive in those areas. The final integrated control command enables combined and differentiated adjustments to key process parameters. This control strategy decouples the overall quality objective from the local defect elimination objective, allowing different actuators to work collaboratively according to different control objectives. This avoids other quality problems that might arise from adjusting a single parameter, improving the precision of control and the overall optimization effect. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the machine learning-based adhesive coating uniformity optimization and control method described in this invention. Figure 2 A flowchart for calculating the predicted value of coating uniformity and identifying potential defect areas for a quality prediction model; Figure 3 Flowchart for establishing and applying control strategy models; Figure 4 Grouped bar chart showing the optimization effect on adhesive coating uniformity; Figure 5 Grouped bar chart comparing control parameters before and after conflict resolution in adhesive coating control. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a machine learning-based method for optimizing and controlling the uniformity of adhesive coating. The method includes: during the adhesive coating production process, the system simultaneously collects data on coating speed, application roller pressure, substrate tension, and adhesive viscosity, forming a real-time process parameter set. Simultaneously, the system receives quality monitoring data from the production line, including adhesive layer thickness distribution maps captured by online image sensors and adhesive layer curing degree distribution maps detected by infrared sensors. The aforementioned real-time process parameter set and quality monitoring data are input into a pre-trained quality prediction model. After calculation, the quality prediction model outputs two key results: a predicted value for the uniformity of the entire coating area (i.e., the coating uniformity prediction value), and coordinates of areas potentially exhibiting defects such as insufficient thickness, uneven curing, or bubbles (i.e., potential defect area identifiers). The system then generates a first control adjustment amount based on the difference between the predicted coating uniformity value and a pre-set target uniformity value. This adjustment amount is specifically used to adjust the coating speed and application roller pressure. Based on the identified potential defect areas, the system generates a second control adjustment amount, which is specifically used to adjust the substrate tension and adhesive viscosity. The system combines the first and second control adjustment amounts to generate a comprehensive control command, which is then sent to the corresponding actuators of the coating equipment. This optimizes the coating process for the next production cycle, achieving and maintaining higher coating uniformity.

[0020] In one embodiment of the present invention, see [reference] Figure 2Taking a coating production line for producing self-adhesive labels as an example, the real-time process parameters include a coating speed set to 30 meters per minute, an application roller pressure set to 200 kPa, a substrate tension set to 150 Newtons, and an adhesive viscosity of 500 mPa·s. The quality monitoring data includes an adhesive layer thickness distribution map captured by an online image sensor and an adhesive layer curing degree distribution map detected by an infrared sensor. The adhesive layer thickness distribution map is presented in the form of a pixel matrix, with each pixel value corresponding to a thickness measurement value at the micrometer level. The adhesive layer curing degree distribution map is presented in the form of a heat map, with each pixel value representing a curing degree measurement value in percentage form. Data comparison shows that when the coating speed increases to 40 meters per minute, the thickness variance of the adhesive layer thickness distribution map increases from 0.5 micrometers squared to 1.2 micrometers squared, while the average curing degree of the adhesive layer curing degree distribution map decreases from 95% to 88%.

[0021] In practical implementation, feature extraction processing is performed on the adhesive layer thickness distribution map to obtain a thickness distribution feature vector. The thickness distribution feature vector includes the average thickness, thickness range, thickness variance, and thickness change trend slope. In the example scenario, the adhesive layer thickness distribution map is divided into 100 uniform grid cells, each containing 1000 pixels. The average thickness of the adhesive layer in each grid cell is calculated by accumulating the thickness values ​​of all pixels in the grid cell and dividing by the number of pixels, 1000. The maximum and minimum thickness values ​​in all grid cells are recorded, and the difference between the maximum and minimum thickness values ​​is used to calculate the thickness range. The average thickness distribution of all grid cells is statistically analyzed, and the standard deviation of the average thickness is calculated as the thickness variance. The average thickness sequence of 10 consecutively arranged grid cells along the coating direction is linearly fitted, and the slope of the fitted line is used as the thickness change trend slope. The linear fitting process uses the least squares method, and the slope calculation formula is: in: The slope representing the trend of thickness variation. This represents the number of grid cells and has a value of 10. This represents the sequential position number of the i-th grid cell along the coating direction. The average thickness of the i-th grid cell is represented by the data comparison, which shows that when the slope of the thickness change trend is positive, the predicted coating uniformity is often lower than the preset target uniformity.

[0022] In some embodiments, feature extraction processing is performed on the adhesive curing degree distribution map to obtain a curing degree distribution feature vector. The curing degree distribution feature vector includes the average curing degree, the curing degree uniformity coefficient, and the coordinates of the curing degree anomaly points. The average curing degree is obtained by summing the curing degree values ​​of all pixels in the adhesive curing degree distribution map and dividing by the total number of pixels. The curing degree uniformity coefficient is obtained by calculating the coefficient of variation of the curing degree values ​​of all pixels. The coordinates of the curing degree anomaly points are obtained by detecting the positions of pixels with curing degree values ​​below 90% of the threshold and recording their row and column indices. In the example scenario, data comparison shows that when the pressure of the adhesive roller is reduced from 200 kPa to 180 kPa, the curing degree uniformity coefficient increases from 0.05 to 0.1, and the number of curing degree anomaly point coordinates increases from 5 to 20.

[0023] It is understandable that the thickness distribution feature vector, the curing degree distribution feature vector, and the real-time process parameter set are normalized and fused to generate a comprehensive feature matrix containing multi-dimensional time series. The normalization process uses the min-max scaling method to scale each feature value to the range of 0 to 1. The fusion process is to concatenate all features into a two-dimensional matrix according to the time step. The rows of the matrix represent time sampling points, and the columns represent feature dimensions. In the example scenario, the comprehensive feature matrix contains data for 10 time steps, and each time step has 12 feature dimensions. Data comparison shows that the unnormalized original feature input will slow down the convergence speed of deep neural network training by 30%.

[0024] Optionally, the comprehensive feature matrix is ​​input into the deep neural network of the quality prediction model. The deep neural network includes a long short-term memory network layer and an attention mechanism layer. The long short-term memory network layer has 50 hidden units to capture the temporal dependency between process parameters and quality data. The attention mechanism layer calculates the weight of the feature at each time step. Features with larger weights contribute more to the output. The deep neural network generates a predicted value for the coating uniformity of the entire coating area and outputs the coordinates of areas with excessively thin thickness, uneven curing, or risk of bubble formation on the adhesive layer thickness distribution map as potential defect area identifiers. In the example scenario, the output layer of the deep neural network uses a linear activation function to generate the predicted value for coating uniformity and a sigmoid activation function to generate a probability map of potential defect area identifiers. Data comparison shows that when the attention mechanism layer is used...

[0025] In some embodiments, the construction steps of the pre-trained quality prediction model include collecting multiple sets of process parameter data and corresponding quality inspection result data during the historical adhesive coating process. The process parameter data includes coating speed, application roller pressure, substrate tension, and adhesive viscosity. The quality inspection result data includes measured values ​​of adhesive layer thickness distribution and adhesive layer curing degree distribution. In the example scenario, the historical data is collected from the production records of the same coating production line over the past year, resulting in 10,000 sets of samples. Each set of samples includes process parameter data and quality inspection result data. Data comparison shows that when the historical data is divided into training and validation sets in a 7:3 ratio, the model has the smallest prediction error on the validation set.

[0026] In practice, historical process parameter data and quality inspection result data are cleaned and labeled to form a training sample set. The cleaning process includes removing samples with more than 10% missing values ​​and correcting outliers that are significantly outside the physical range. The labeling process assigns a unique identifier to each sample and associates it with the process parameter data and quality inspection result data. In the example scenario, 9,500 samples are retained after data cleaning. The labeling process ensures that the process parameter data and quality inspection result data of each sample correspond one-to-one. Data comparison shows that the cleaned training sample set improves the model training stability by 20%.

[0027] Optionally, the deep neural network model is trained using a training sample set, and the model parameters are optimized using the backpropagation algorithm until the error between the model's predicted output and the quality detection results is lower than a preset threshold. The training process uses mean squared error as the loss function and the Adam optimizer is used to update the parameters. The preset threshold is set to 0.01. In the example scenario, after 500 training iterations, the loss function value drops to 0.009, reaching the preset threshold. Data comparison shows that using the Adam optimizer is 40% faster than using the stochastic gradient descent optimizer.

[0028] In practice, the trained deep neural network model is saved as a pre-trained quality prediction model, which is used to receive real-time process parameter sets and quality monitoring data and output prediction results in real time. The saved format is HDF5 file, which includes the model architecture and weight parameters. In the example scenario, the pre-trained quality prediction model is deployed on the edge computing device of the coating production line. The real-time inference latency is less than 100 milliseconds. Data comparison shows that using the pre-trained model saves 95% of the computation time compared to retraining the model each time.

[0029] In one embodiment of the present invention, on a precision coating equipment for producing optical thin films, an online image sensor captures an adhesive layer thickness distribution map with a spatial resolution of 1024 pixels by 1024 pixels, covering a coating area with a width of 1 meter. The thickness value of each pixel in the adhesive layer thickness distribution map is stored in floating-point form in micrometers. Data comparison shows that when the temperature of the coating roller increases from 50 degrees Celsius to 60 degrees Celsius, the overall grayscale distribution of the adhesive layer thickness distribution map changes significantly, and the average value of the region with the maximum thickness decreases from 25.3 micrometers to 24.1 micrometers. In practical implementation, the adhesive layer thickness distribution map is divided into multiple uniform grid units. The size of the grid unit is determined according to the coating width and accuracy requirements. In the example scenario, the 1024-pixel by 1024-pixel adhesive layer thickness distribution map is uniformly divided into 32 parts along both the width and length directions, thereby generating 1024 uniform grid units. Each grid unit contains 32-pixel by 32-pixel points, totaling 1024 pixels. The division process is implemented through an image processing algorithm. The algorithm assigns a unique row and column index coordinate to each grid unit. Data comparison shows that when the number of grid units is increased from 256 to 1024, the sensitivity of the slope of the thickness change trend obtained in subsequent calculations to local thickness fluctuations is improved.

[0030] It is understandable that the average thickness of the adhesive layer in each grid cell is calculated, and the maximum and minimum thickness values ​​are recorded. The thickness range is calculated by the difference between the maximum and minimum thickness values. When calculating the average thickness of the adhesive layer in each grid cell, the algorithm traverses all 1024 pixels in the grid cell, sums the thickness values ​​of each pixel, and divides by 1024 to obtain the average thickness of the adhesive layer in that grid cell. During the traversal of all 1024 grid cells, the maximum and minimum values ​​of the average thickness of the adhesive layer in all grid cells are recorded and updated simultaneously. The arithmetic difference between the maximum and minimum thickness values ​​is the thickness range. In the example scenario, when the substrate tension fluctuates by 5 Newtons, the thickness range increases from 2.1 micrometers to 3.8 micrometers.

[0031] In some embodiments, the average thickness distribution of all grid cells is statistically analyzed, and the standard deviation of the average thickness is calculated as the thickness variance. The statistical process involves collecting the average thickness values ​​of the adhesive layer from all 1024 grid cells, forming a sample set containing 1024 data points, and calculating the standard deviation of this data set as the thickness variance. The formula for calculating the standard deviation is as follows: in: Represents the thickness variance. This represents the total number of grid cells and has a value of 1024. The average thickness of the adhesive layer represents the k-th grid cell. The arithmetic mean of the average thickness of the adhesive layer for all grid cells is represented by the data comparison, which shows that when the adhesive viscosity is adjusted from 800 mPa·s to 750 mPa·s, the calculated thickness variance decreases from 1.05 μm² to 0.82 μm².

[0032] Optionally, the average thickness sequence of multiple consecutively arranged grid cells is linearly fitted along the coating direction, and the slope of the fitted line is used as the slope of the thickness change trend. In the example scenario, eight consecutively arranged grid cells in the same column are selected along the traveling direction of the coating machine, the average thickness value of the adhesive layer of these eight grid cells is extracted and arranged in spatial order to form an average thickness sequence, and the sequence is fitted using a linear regression method. The slope parameter of the fitted line is extracted as the slope of the thickness change trend. This slope value is used to characterize the linear increasing or decreasing trend of the adhesive layer thickness along the coating direction.

[0033] In one embodiment of the present invention, see [reference] Figure 3 The process involves generating a first control adjustment based on the difference between the predicted coating uniformity and the preset target uniformity. On a lithium battery separator coating production line, the current predicted coating uniformity value output by the quality prediction model is 0.89, while the preset target uniformity in the production procedure is 0.95. The system calculates an absolute difference of 0.06. Data comparison shows that when the absolute difference is 0.02, the change in process parameters caused by the first control adjustment is smaller than when the absolute difference is 0.08. In specific implementation, the absolute difference between the predicted coating uniformity and the preset target uniformity is calculated. The absolute difference is the absolute value of the result after subtracting the preset target uniformity from the predicted coating uniformity. In the example scenario, the predicted coating uniformity comes from the output of the quality prediction model's calculation of the comprehensive feature matrix. The preset target uniformity is set to a fixed value of 0.95 by the process personnel according to the product specifications. The calculated absolute difference is 0.06, which is transmitted to the control strategy model in real time. Data comparison shows that when the absolute difference is greater than 0.05 for more than three control cycles, the system will trigger an additional alarm.

[0034] It is understandable that the absolute difference is input into the control strategy model, which contains a nonlinear relationship function between coating speed and sizing roller pressure. The control strategy model uses the absolute difference as the main input variable. The model has a pre-stored mapping table of the coupling relationship between the coating speed adjustment and the sizing roller pressure adjustment. In the example scenario, after the absolute difference of 0.06 is input into the control strategy model, the model queries the internal nonlinear relationship function and outputs a set of preliminary adjustment parameters. Data comparison shows that when the input absolute difference is the same, the adjustment amount calculated by the control strategy model through the nonlinear relationship function differs for adhesives of different viscosities.

[0035] In some embodiments, the control strategy model calculates that when the absolute difference exceeds a first threshold, it generates a first speed adjustment amount to reduce the coating speed and a first pressure adjustment amount to increase the pressure of the sizing roller. In the example scenario, the first threshold is set to 0.05. Since the absolute difference of 0.06 is greater than the first threshold of 0.05, the control strategy model generates a first speed adjustment amount to reduce the coating speed according to its internal calculation rules. The first speed adjustment amount suggests reducing the current coating speed from 25 meters per minute to 23.5 meters per minute. At the same time, it generates a first pressure adjustment amount to increase the pressure of the sizing roller. The first pressure adjustment amount suggests increasing the current sizing roller pressure from 120 kPa to 125 kPa. Data comparison shows that when the first threshold is adjusted from 0.05 to 0.04, the system intervenes more frequently in adjusting the speed and pressure.

[0036] Optionally, the control strategy model calculates that when the absolute difference is lower than the second threshold, it generates a second speed adjustment amount to increase the coating speed and a second pressure adjustment amount to decrease the pressure of the sizing roller. In another example scenario, if the predicted coating uniformity is 0.97 and the preset target uniformity is 0.95, the absolute difference is 0.02, and the second threshold is set to 0.03. Since the absolute difference of 0.02 is lower than the second threshold of 0.03, the control strategy model generates a second speed adjustment amount to increase the coating speed, which suggests increasing the coating speed from 25 meters per minute to 26 meters per minute. At the same time, it generates a second pressure adjustment amount to decrease the pressure of the sizing roller, which suggests decreasing the pressure of the sizing roller from 120 kPa to 118 kPa. Data comparison shows that the setting of the second threshold affects production efficiency; a lower threshold will lead to more equipment adjustment actions.

[0037] In specific implementation, the first speed adjustment amount is combined with the first pressure adjustment amount, or the second speed adjustment amount is combined with the second pressure adjustment amount to form the first control adjustment amount. In the example where the absolute difference exceeds the first threshold, the generated first speed adjustment amount and the first pressure adjustment amount are encapsulated into a structured data packet, which constitutes the first control adjustment amount. The first control adjustment amount contains two fields that store the speed adjustment command and the pressure adjustment command respectively. In the example where the absolute difference is lower than the second threshold, the generated second speed adjustment amount and the second pressure adjustment amount are combined in the same way. Data comparison shows that the combined first control adjustment amount is sent to the actuator through the industrial network in data frame format.

[0038] It is understandable that a dynamic relationship model is established between coating speed and application roller pressure. This model defines the mapping relationship between the change in coating speed and the required change in application roller pressure when maintaining a constant coating amount per unit area under a specific adhesive viscosity. The basic form of the dynamic relationship model is a function: in: This represents the amount of change in the required application roller pressure. This represents the change in coating speed. Representing the current viscosity of the adhesive, in the example scenario, for an adhesive with a viscosity of 550 mPa·s, the dynamic relationship model shows that for every 1 meter per minute increase in coating speed, the application roller pressure needs to be reduced by approximately 2.2 kPa to maintain the same coating amount. Data comparison shows that for different substrates, the function... The coefficients in the code need to be recalibrated.

[0039] In some embodiments, the absolute difference is introduced as a feedback factor in the dynamic relationship model. The sign and magnitude of the absolute difference are used to dynamically adjust the scaling factor of the mapping relationship. The feedback factor is weighted by a coefficient. The form is introduced, and the modified relationship is: The weighting coefficients and absolute difference The relationship is , The sensitivity coefficient set for the system, in the example scenario, is... Set to 2, when the absolute difference When the weighting coefficient is 0.06, The calculated value is 1.12, which means that based on the absolute difference feedback, the adjustment amount of the base pressure calculated by the dynamic relationship model is amplified by 12%. Data comparison shows that the sensitivity coefficient... The value of directly affects the system's response strength to uniformity deviation.

[0040] Optionally, a control strategy model containing corrected proportional coefficients is obtained by training the dynamic relationship model using offline training data. This model is used to accurately calculate the adjustment amount of speed and pressure based on the absolute difference of the real-time input. The offline training data comes from coating speed, application roller pressure, adhesive viscosity, and final measured coating uniformity data stored in the historical production database. The training process adjusts the function through an optimization algorithm. Parameters and sensitivity coefficients in This allows the model's predicted adjustment amount to correct uniformity deviation to the greatest extent possible. The trained control strategy model encapsulates the corrected proportional coefficient and calculation logic.

[0041] In one embodiment of the present invention, during the coating process of a roll of release paper with a width of 1200 mm, the quality prediction model outputs three potential defect area markers with coordinate ranges of 200 mm to 250 mm in the width direction, 600 mm to 620 mm in the width direction, and 950 mm to 1000 mm in the width direction, respectively. The system parses these markers and initiates the generation logic of the second control adjustment amount. Data comparison shows that when the number of potential defect area markers increases from one to three, the calculation time for generating the second control adjustment amount is correspondingly extended.

[0042] In practical implementation, the potential defect area identifiers are parsed to identify the location distribution and defect type of the potential defect areas in the coating width direction. The parsing process reads the coordinate information and classification probability contained in the potential defect area identifiers. The location distribution is determined by the start and end pixel positions in the coordinate information and converted into actual millimeter units. The defect type is determined by the classification labels output by the quality prediction model. The labels include "too thin", "uneven curing", and "bubbles". In the example scenario, the parsing results identify the defect type as "too thin" at a width of 200 mm to 250 mm, "uneven curing" at a width of 600 mm to 620 mm, and "bubbles" at a width of 950 mm to 1000 mm. Data comparison shows that the parsing and response speed requirements for "bubble" type defects are higher than those for other types of defects.

[0043] It is understandable that when the defect type is "too thin", the system calculates the substrate tension required at the corresponding location of the potential defect area and generates an adjustment amount to increase the substrate tension. In the example scenario, for the identified "too thin" defect area, the system calls the tension calculation module. Based on the center coordinates and width of the defect area, combined with the overall tension value of the current substrate, the module calculates a specific value of the substrate tension that needs to be increased. This value is then formatted as a torque control command for a specific tension roller, forming an adjustment amount to increase the substrate tension. Data comparison shows that for a 50 mm wide area of ​​"too thin", the calculated substrate tension required is about 150% higher than that for an area with a 20 mm wide area.

[0044] In some embodiments, when the defect type is uneven curing or bubbles, the system queries the correlation database of adhesive viscosity and curing characteristics, calculates the target adhesive viscosity value required to improve the curing state of the potential defect area, and generates a viscosity adjustment command. The correlation database of adhesive viscosity and curing characteristics is stored in tabular form, recording parameters such as curing rate and final degree of curing for different adhesive formulations at different viscosities. In the example scenario, for the "uneven curing" defect area, the system queries the correlation database and, based on the current degree of curing distribution, oven temperature, and linear velocity, performs a reverse query to obtain the target adhesive viscosity value required to improve curing uniformity. The difference between the target viscosity value and the current viscosity is converted into an adjustment command for the viscosity control valve of the adhesive supply system, i.e., a viscosity adjustment command. Data comparison shows that the average processing time for querying the correlation database and generating the viscosity adjustment command is 15 milliseconds.

[0045] Optionally, the substrate tension increase adjustment amount and viscosity adjustment instruction can be combined to form a second control adjustment amount. In the example scenario, the substrate tension increase adjustment amount and viscosity adjustment instruction generated by the system are encapsulated into the same control data packet. The control data packet has a predefined data structure, including instruction type, target actuator address, adjustment parameter value and timestamp field. This control data packet is used as the second control adjustment amount and is transmitted to the real-time controller of the coating equipment. Data comparison shows that the combined second control adjustment amount data packet size is fixed at 64 bytes.

[0046] In practical implementation, the specific process of calculating the substrate tension required at the corresponding location of the potential defect area and generating the substrate tension increase adjustment amount is as follows: determine the center coordinates and area width of the potential defect area in the coating width direction. In the example, for a defect area with a width of 200 mm to 250 mm, the center coordinate is calculated to be 225 mm and the area width is 50 mm. Data comparison shows that the calculation accuracy of the center coordinate directly affects the accuracy of local tension compensation.

[0047] It is understandable that, based on the center coordinates and the area width, the local tension compensation coefficient is calculated. This local tension compensation coefficient is directly proportional to the area width. The formula is: in: Represents the local tension compensation coefficient. This is a proportionality constant, determined by the characteristics of the base material and the equipment. The width of the region representing the potential defect area, in the example, is a scaling constant. Set to 0.02 per millimeter, area width The local tension compensation coefficient is calculated to be 50 mm. The value is 1.0. Data comparison shows that when the proportionality constant is 1.0... When the value is adjusted from 0.02 to 0.015, the local tension compensation coefficient decreases accordingly.

[0048] In some embodiments, the current overall tension value of the substrate is obtained, and multiplied by a local tension compensation coefficient to obtain the required additional tension of the substrate. The current overall tension value of the substrate is obtained from the real-time reading of the tension sensor, for example, 300 Newtons. The overall tension value of 300 Newtons is multiplied by a local tension compensation coefficient of 1.0 to obtain a required additional tension of 300 Newtons. Data comparison shows that the measurement update frequency of the overall tension value of the substrate needs to be higher than 10 Hz to ensure the real-time performance of the calculation. Optionally, the required additional tension of the substrate is converted into a torque control increment for a specific tension roller, forming an adjustment amount for increasing the substrate tension. The conversion process is based on mechanical parameters such as the radius and transmission ratio of the tension roller, converting the tension value into a motor torque command. In the example, the required additional tension of the substrate is 300 Newtons, and combined with the radius of the tension roller of 0.1 meters, the calculated torque increase is 30 N·m. This torque control increment, together with the address code of the executing motor, constitutes the core content of the adjustment amount for increasing the substrate tension. Data comparison shows that the delay from the completion of the calculation to the issuance of the torque command to the driver is less than 5 milliseconds. See Table 1.

[0049] Table 1: Database Lookup Table for the Correlation between Adhesive Viscosity and Curing Properties See Figure 4 This is a grouped bar chart showing the effect of adhesive coating uniformity optimization. It compares coating quality indicators "before optimization" and "after optimization," verifying the effectiveness of control command execution. The optimization magnitude varies across different indicators, with some (such as the last one) showing significantly greater improvement than others, indicating that the control strategy is more targeted at specific quality dimensions. It quantifies the actual effect of coating uniformity optimization control methods, visually demonstrating the improvement in core quality indicators such as thickness and cure degree after process parameter adjustments, and is a key basis for verifying the rationality of the control strategy. In adhesive coating production, this type of chart is a core tool demonstrating the value of "machine learning-based optimization control," assisting manufacturers in evaluating the actual benefits of process improvements.

[0050] In one embodiment of the present invention, on a composite film coating production line, the first control adjustment includes an instruction to reduce the coating speed from 30 meters per minute to 28 meters per minute and an instruction to increase the pressure of the application roller from 200 kPa to 205 kPa. The second control adjustment includes an instruction to increase the substrate tension from 100 Newtons to 110 Newtons and an instruction to adjust the adhesive viscosity from 450 mPa·s to 440 mPa·s. The system receives the above two adjustment quantities and starts the process of generating comprehensive control instructions. Data comparison shows that in 100 consecutive control cycles, the execution conflict between the first control adjustment and the second control adjustment occurs on average 5 times. In practical implementation, the system determines whether there is an execution conflict between the first control adjustment amount and the second control adjustment amount. The judgment logic is based on a predefined conflict rule base. The rule base defines the mutual exclusion relationship between the actions of different actuators. In the example scenario, the system detects that the first control adjustment amount requires a reduction in coating speed. According to the coating process model, reducing the coating speed usually requires a simultaneous reduction in substrate tension to maintain stability. However, the second control adjustment amount requires an increase in substrate tension. This set of instructions is judged by the conflict rule base as having an execution conflict. Data comparison shows that the reverse adjustment combination of coating speed and substrate tension accounts for more than 60% of the conflict records.

[0051] It is understandable that when there is no execution conflict, the first control adjustment amount and the second control adjustment amount are directly merged to generate a comprehensive control instruction. In another example scenario, the first control adjustment amount includes instructions to increase the coating speed and reduce the pressure of the application roller, while the second control adjustment amount only includes instructions to increase the adhesive viscosity. The conflict rule base determines that these two sets of instructions act on different and non-interfering actuators. Therefore, the system packages the speed adjustment instruction, pressure adjustment instruction, and viscosity adjustment instruction in chronological order to form a complete comprehensive control instruction data packet and sends it.

[0052] In some embodiments, when execution conflicts exist, the parameters in the first control adjustment amount or the second control adjustment amount are adjusted according to a preset conflict resolution rule. The conflict resolution rule prioritizes ensuring coating thickness uniformity and then considers curing uniformity. In an example where a conflict exists, the conflict resolution rule determines that the influence of coating speed on thickness uniformity has a higher weight than substrate tension. Therefore, the instruction to reduce coating speed in the first control adjustment amount is retained. At the same time, the system modifies the instruction to increase substrate tension in the second control adjustment amount according to a compensation algorithm. The modified instruction may reduce the tension increase from 10 Newtons to 3 Newtons or delay execution. Data comparison shows that after applying the conflict resolution rule, the phenomenon of increased coating thickness standard deviation caused by parameter conflict is reduced by 90%.

[0053] Optionally, the conflict resolution rules are adjusted using an optimization function, which takes the following form: in: The expected total improvement in coating uniformity is determined by the difference between the predicted coating uniformity and the preset target uniformity. This represents the amount of adjustment for the coating speed; This represents the adjustment amount of the application roller pressure. This represents the amount of substrate tension adjustment; , , These represent the weighting coefficients for the effects of coating speed, application roller pressure, and substrate tension on thickness uniformity, respectively. In the event of a conflict, the system remains fixed. and This function is used to find a new value that minimizes conflict with the first control adjustment. Replace the value in the original second control adjustment.

[0054] In practice, the adjusted control values ​​are merged to generate the final integrated control command. After the conflict is resolved, the system recombines the modified second control value with the unmodified first control value. The combination process ensures that the timestamps of all commands are synchronized and the execution order is logically correct. The final result is a structured command set, which is encapsulated in a protocol format that the device controller can recognize. In the example scenario, the final integrated control command contains four sub-commands: setting the coating speed to 28 meters per minute, setting the application roller pressure to 205 kPa, setting the substrate tension to 103 Newtons, and setting the adhesive viscosity to 440 mPa·s. Data comparison shows that the average time from conflict detection to generating the final command is less than 50 milliseconds.

[0055] See Figure 5 This is a grouped bar chart comparing control parameters before and after conflict resolution in adhesive coating control. It displays the numerical comparison of four control parameters—coating speed, application roller pressure, substrate tension, and adhesive viscosity—under "baseline value, initial adjustment, and post-resolution adjustment," corresponding to the parameter optimization results in the conflict resolution stage. It visually demonstrates the effectiveness of conflict resolution strategies: by fine-tuning mutually exclusive parameters (such as substrate tension), it eliminates execution conflicts of control commands while ensuring coating thickness uniformity, serving as a key verification tool for ensuring coating process stability. In the control process of adhesive coating production, this type of chart is the core basis for demonstrating the rationality of "conflict resolution rules," assisting process engineers in evaluating the actual implementation effect of control strategies.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing control of adhesive coating uniformity based on machine learning, characterized by, The method comprises: collecting a set of real-time process parameters during the adhesive coating process, the set of real-time process parameters including coating speed, glue roller pressure, substrate tension, and viscosity data of the adhesive; receiving quality monitoring data of the coating area, the quality monitoring data including a glue layer thickness distribution map captured by an online image sensor and a glue layer curing degree distribution map detected by an infrared sensor; inputting the set of real-time process parameters and the quality monitoring data into a pre-trained quality prediction model, and outputting a coating uniformity prediction value and a potential defect area identification through calculation of the quality prediction model; generating a first control adjustment amount based on the difference between the coating uniformity prediction value and a preset target uniformity, the first control adjustment amount being used to adjust the coating speed and the glue roller pressure; generating a second control adjustment amount based on the potential defect area identification, the second control adjustment amount being used to adjust the substrate tension and the viscosity of the adhesive; generating a comprehensive control instruction based on the first control adjustment amount and the second control adjustment amount, and sending the comprehensive control instruction to a coating device for execution to control the coating process of the next production cycle. 2.The machine learning-based adhesive coating uniformity optimization control method of claim 1, wherein The inputting of the set of real-time process parameters and the quality monitoring data into a pre-trained quality prediction model, and the outputting of a coating uniformity prediction value and a potential defect area identification through calculation of the quality prediction model, comprises: performing feature extraction processing on the glue layer thickness distribution map to obtain a thickness distribution feature vector, the thickness distribution feature vector including average thickness, thickness range, thickness variance, and thickness change trend slope; performing feature extraction processing on the glue layer curing degree distribution map to obtain a curing degree distribution feature vector, the curing degree distribution feature vector including average curing degree, curing degree uniformity coefficient, and curing degree abnormal point coordinates; performing normalized fusion processing on the thickness distribution feature vector, the curing degree distribution feature vector, and the set of real-time process parameters to generate a comprehensive feature matrix including multi-dimensional time series; inputting the comprehensive feature matrix into a deep neural network of the quality prediction model, the deep neural network including a long short-term memory network layer and an attention mechanism layer; generating a coating uniformity prediction value for the entire coating area through calculation of the deep neural network, and outputting the coordinate positions of the areas at risk of having excessively thin thickness, uneven curing, or generating bubbles on the glue layer thickness distribution map as the potential defect area identification. 3.The machine learning-based adhesive coating uniformity optimization control method of claim 2, wherein The feature extraction processing on the glue layer thickness distribution map to obtain a thickness distribution feature vector comprises: dividing the glue layer thickness distribution map into a plurality of uniform grid cells; calculating the average thickness of each grid cell, and recording the maximum thickness and the minimum thickness, and calculating the thickness range from the difference between the maximum thickness and the minimum thickness; statistically analyzing the average thickness distribution of all grid cells, and calculating the standard deviation of the average thickness as the thickness variance; linearly fitting the average thickness sequence of the plurality of grid cells arranged in series along the coating direction, and taking the slope of the fitted straight line as the thickness change trend slope. 4.The machine learning-based adhesive coating uniformity optimization control method of claim 1, wherein The first control adjustment amount is generated according to a difference between the coating uniformity prediction value and a preset target uniformity, and the method comprises the following steps: calculating an absolute difference between the coating uniformity prediction value and the preset target uniformity; inputting the absolute difference into a control strategy model, wherein the control strategy model comprises a nonlinear relationship function between the coating speed and the size of the glue roller pressure; calculating, through the control strategy model, that when the absolute difference exceeds a first threshold value, a first speed adjustment amount for reducing the coating speed and a first pressure adjustment amount for increasing the glue roller pressure are generated; when the absolute difference is less than a second threshold value, a second speed adjustment amount for increasing the coating speed and a second pressure adjustment amount for reducing the glue roller pressure are generated; combining the first speed adjustment amount and the first pressure adjustment amount, or combining the second speed adjustment amount and the second pressure adjustment amount, to form the first control adjustment amount. 5.The machine learning-based adhesive coating uniformity optimization control method of claim 4, wherein The control strategy model comprises a nonlinear relationship function between the coating speed and the size of the glue roller pressure, and the method comprises the following steps: establishing a dynamic relationship model between the coating speed and the size of the glue roller pressure, wherein the dynamic relationship model defines a mapping relationship between the change amount of the coating speed and the change amount of the required glue roller pressure when maintaining a constant unit area coating amount under a specific adhesive viscosity; introducing the absolute difference as a feedback factor in the dynamic relationship model, and dynamically adjusting the proportional coefficient of the mapping relationship through the positive and negative and size of the absolute difference; training the dynamic relationship model through offline training data to obtain a control strategy model comprising a corrected proportional coefficient, which is used to accurately calculate the adjustment amount of the speed and the pressure according to the real-time input absolute difference. 6.The machine learning-based adhesive coating uniformity optimization control method of claim 1, wherein The second control adjustment amount is generated according to the potential defect area identification, and the method comprises the following steps: analyzing the potential defect area identification to identify the position distribution of the potential defect area in the coating width direction and the defect type; when the defect type is too thin, calculating the required additional tension of the substrate at the position corresponding to the potential defect area to generate a substrate tension increase adjustment amount; when the defect type is curing unevenness or bubble, querying an associated database of adhesive viscosity and curing characteristics, calculating the target value of the adhesive viscosity required to improve the curing state of the potential defect area, and generating a viscosity adjustment instruction; combining the substrate tension increase adjustment amount and the viscosity adjustment instruction to form the second control adjustment amount. 7.The machine learning-based adhesive coating uniformity optimization control method of claim 6, wherein The method for calculating the required additional tension of the substrate at the position corresponding to the potential defect area to generate a substrate tension increase adjustment amount comprises the following steps: determining the center coordinates and the area width of the potential defect area in the coating width direction; calculating a local tension compensation coefficient according to the center coordinates and the area width, wherein the local tension compensation coefficient is proportional to the area width; obtaining the current overall tension value of the substrate, multiplying the overall tension value of the substrate by the local tension compensation coefficient to obtain the required additional tension of the substrate; and converting the required additional tension of the substrate into a torque control increment of a specific tension roller to form the substrate tension increase adjustment amount. 8.The machine learning-based adhesive coating uniformity optimization control method of claim 1, wherein The generating a comprehensive control instruction based on the first control adjustment amount and the second control adjustment amount comprises: determining whether the first control adjustment amount and the second control adjustment amount have an execution conflict; when there is no execution conflict, directly combining the first control adjustment amount and the second control adjustment amount to generate the comprehensive control instruction; when there is an execution conflict, adjusting parameters in the first control adjustment amount or the second control adjustment amount according to a preset conflict resolution rule, the conflict resolution rule giving priority to guaranteeing coating thickness uniformity and then considering curing uniformity; combining the adjusted control adjustment amount to generate a final comprehensive control instruction. 9.The machine learning-based adhesive coating uniformity optimization control method of claim 1, wherein, The construction steps of the pre-trained quality prediction model comprise: collecting a plurality of sets of historical adhesive coating process parameter data and corresponding quality detection result data, the process parameter data including coating speed, glue roller pressure, substrate tension, and adhesive viscosity, and the quality detection result data including glue layer thickness distribution measurement values and glue layer curing degree distribution measurement values; cleaning and labeling the historical process parameter data and quality detection result data to form a training sample set; constructing a deep neural network model as a basic structure of the quality prediction model, the deep neural network model including an input layer, a plurality of hidden layers, and an output layer; training the deep neural network model using the training sample set, optimizing model parameters through a back propagation algorithm until the error between the model prediction output and the quality detection result data is below a preset threshold; saving the trained deep neural network model as the pre-trained quality prediction model for real-time receiving of the real-time process parameter set and the quality monitoring data and outputting a prediction result. 10.A machine learning based adhesive coating uniformity optimization control system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the adhesive coating uniformity optimization control method based on machine learning in any one of claims 1 to 9.

Images