A production rubber roll coating device and method based on wet paste

Abstract

The application discloses a kind of production rubber roller film coating device and method based on wet paste, belong to plastic laminating forming technical field, it is related to the composite forming process of plastic film and carbon fiber cloth.Device includes frame body, frame body is equipped with glue tank, tank is filled with macromolecular adhesive, inside is provided with rubber roller;Frame body outside is equipped with driving motor, for driving rubber roller rotation;Frame body inner wall is located just above rubber roller rotation connection guide roller.The scraper of frame body inner guide roller is set, for the macromolecular adhesive of carbon fiber cloth surface is scraped, frame body is close to the servo motor of scraper, controls scraper movement, and pressure sensor is installed on scraper.Glue tank is equipped with viscosity sensor, detects macromolecular adhesive viscosity;Frame body is equipped with camera, obtains the image data of carbon fiber cloth surface after scraping glue.The application improves plastic laminating uniformity, reduces glue layer cracking or bonding deficiency and other quality problems, and is suitable for efficient production in plastic forming field.

Description

Technical Field

[0001] This invention relates to the field of plastic lamination technology, and more specifically, to a production roller coating apparatus and method based on wet paste preparation. Background Technology

[0002] In the production of carbon fiber membranes, a wet lamination process is often used for coating. Generally, the composite material is first fed and replaced using an automatic feeding and roll-to-roll conversion mechanism. The composite material can be carbon fiber cloth. Before lamination, the carbon fiber cloth needs to undergo a wet impregnation process. During impregnation, the impregnation tank is equipped with an upper and lower roller, the lower half of which is completely immersed in a glue bath containing polymer adhesive. When the carbon fiber cloth enters the impregnation stage, it slowly passes between the upper and lower rollers. The lower roller, driven by a drive device, begins to rotate. Because its lower half is immersed in the glue bath, and the lower roller is typically a wool roller or bristle roller, its rotation evenly spreads the polymer adhesive from the glue bath onto the lower surface of the carbon fiber cloth. After the polymer adhesive is applied, a layer of the required plastic film is immediately covered on the adhesive-covered side. At this point, the polymer adhesive is in a undried state, and this moist state provides favorable conditions for good adhesion between the plastic film and the carbon fiber cloth. Subsequently, pressure rollers are used to apply pressure to the covered plastic film and carbon fiber cloth, ensuring a tight bond between them using adhesive. This prevents the polymer adhesive from contaminating the cutting machine during subsequent cutting. The coated carbon fiber cloth can be cut to any size according to the mold requirements. For example, if the mold is for a drone shell, it can be disassembled as needed. The disassembled carbon fiber cloth is then placed between the convex and concave molds and joined together, laying the carbon fiber cloth flat within the mold. The mold is then placed in an autoclave under a relative vacuum. Through heating, pressurizing, heat preservation, cooling, and depressurization processes, the carbon fiber cloth is shaped to the desired form and quality. The autoclave is then opened, the mold removed, and the processed carbon fiber drone shell can be taken out. Further deburring of the carbon fiber drone shell yields a qualified carbon fiber drone shell.

[0003] In the aforementioned impregnation process, the lower roller continuously applies polymeric adhesive to the lower surface of the carbon fiber cloth, gradually ensuring the cloth is fully saturated. During this process, the amount and degree of saturation of the adhesive must be strictly controlled. Insufficient adhesive application may result in weak adhesion between the carbon fiber cloth and the subsequent laminating film, affecting the strength and durability of the laminated product. Conversely, excessive adhesive application not only wastes the adhesive and increases production costs but may also lead to adhesive dripping and accumulation during subsequent processing, impacting the product's appearance and performance.

[0004] After impregnation, to ensure uniform adhesive thickness on the carbon fiber cloth surface, existing technologies use a scraper to remove excess adhesive from the impregnated carbon fiber cloth. The scraper contacts the carbon fiber cloth surface at a specific angle and pressure. However, although the scraper can control the adhesive thickness to some extent, some problems still exist in practice. For example, the adhesive coating may be too thick or too thin. An excessively thick coating leads to prolonged drying or curing time, increasing production costs, and may cause problems such as adhesive layer cracking during product use. An excessively thin coating affects the bond strength between the carbon fiber cloth and the film, reducing the overall performance of the product, and may also result in uneven adhesive distribution, localized missing adhesive, and other quality issues, all of which adversely affect the final coated product. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a production roller coating device and method based on wet paste preparation. This device and method can monitor and adjust the scraping parameters in real time through multiple sensors, accurately calculate the thickness deviation of the adhesive layer using an AI model, and generate optimization suggestions by combining pressure and viscosity data. This significantly improves the uniformity of coating, reduces quality problems such as adhesive layer cracking or insufficient adhesion, and enhances the overall performance of the product.

[0006] To solve the above problems, the present invention adopts the following technical solution.

[0007] A production roller coating device based on wet paste forming includes a frame, an adhesive tank installed on the frame, an adhesive roller disposed inside the adhesive tank, a drive motor installed on the outer side of the frame for driving the adhesive roller to rotate, and a guide roller rotatably connected to the inner wall of the frame directly above the adhesive roller.

[0008] A scraper blade is positioned above the guide roller within the frame. This scraper blade is used to apply polymer adhesive to the surface of the carbon fiber cloth. A servo motor is also installed on the frame near the scraper blade, controlling its movement. A pressure sensor is mounted on the scraper blade to detect the pressure applied by the scraper blade to the carbon fiber cloth. A viscosity sensor is installed in the adhesive tank to detect the viscosity of the polymer adhesive within the tank. A camera is also mounted on the frame to acquire image data of the carbon fiber cloth surface after the scraper blade has applied the adhesive. Using the initial scraping image data and polymer adhesive viscosity data acquired by the camera, combined with the real-time contact pressure value of the scraper blade monitored by the pressure sensor, an AI segmentation model is used to analyze the image and calculate the adhesive layer thickness and its deviation. A learning decision model, based on the adhesive layer thickness deviation, current contact pressure value, and polymer adhesive viscosity data, generates optimized scraping pressure adjustment suggestions and safety constraint commands, outputting the target displacement value of the servo motor and the target pressure value of the scraper blade. The servo motor then dynamically adjusts the scraper blade movement accordingly.

[0009] Furthermore, the linkage parameters between the initial thickness of the polymer adhesive on the carbon fiber cloth and the state of the polymer adhesive are set in advance, and the pressure monitoring threshold of the scraper is set. The contact pressure is collected in real time through the pressure sensor, and the mapping relationship between the initial thickness and the corresponding pressure value is recorded simultaneously.

[0010] The adhesive layer after initial coating is continuously photographed by a camera. Based on the detected pressure anomalies, the sampling point density is increased at the corresponding positions in the image. The image pixels are bound to the actual position and pressure value of the carbon fiber cloth through a spatiotemporal alignment algorithm, generating a mapping table of image pixels, actual coordinates and pressure values.

[0011] Furthermore, the image is segmented into adhesive regions using an AI segmentation model, and the actual thickness of each region is calculated through grayscale value analysis. The pressure data at the corresponding locations is matched using a coordinate mapping table to distinguish the correlation between thicker and thinner thicknesses. At the same time, a three-dimensional correlation model of thickness deviation, scraping pressure, and polymer adhesive viscosity is established based on the state data of the polymer adhesive. Based on the three-dimensional correlation model, pressure adjustment suggestions for the scraper are derived.

[0012] Furthermore, a learning decision model is established. This model receives thickness deviation data from image analysis and pressure adjustment suggestions from the scraper. The thickness uniformity improvement rate is used as the core reward function. Combined with safety constraints, the optimal instructions are generated, and the target values ​​for the displacement of the servo motor and the pressure of the scraper are given.

[0013] Furthermore, the glue scraper is driven to move according to the target displacement value; the movement trajectory of the glue scraper is recorded synchronously; if the trajectory fluctuates, feedback is provided in real time and the control of the servo motor is smoothly corrected.

[0014] Furthermore, based on the pressure monitored and adjusted by the pressure sensor, the deviation between the pressure and the target pressure value is calculated, and the deviation value is compared with a preset deviation judgment threshold. If the absolute value of the deviation value is greater than the deviation judgment threshold, it is fed back to the servo motor in real time, and the servo motor controls the scraper to move backward, while the pressure stabilization time is recorded.

[0015] Furthermore, a database of pressure, images, adjustment trajectories, and thickness was established. The model was trained and optimized through machine learning, and the AI ​​segmentation model was fine-tuned using new defect samples.

[0016] Furthermore, a support plate is fixedly installed at the lower end of the camera, and limit plates are provided at both the upper and lower ends of the glue scraper. The front and rear ends of the limit plates are fixedly connected to the inner wall of the frame. An installation plate is provided on the left side of the glue scraper, and a pressure sensor is disposed between the glue scraper and the installation plate. The pressure sensor is fixedly connected to the installation plate, and an elastic block is fixedly connected between the glue scraper and the installation plate. A fixing block is fixedly installed on the left side of the installation plate, and a lead screw is threaded onto the fixing block. The left end of the lead screw is fixedly connected to the output end of the servo motor. A support plate is fixedly connected to the lower end of the servo motor, and the front and rear ends of the support plate and the support plate are fixedly connected to the inner wall of the frame.

[0017] Furthermore, the drive motor is fixedly installed on the outer side of the frame, the output shaft of the drive motor is fixedly connected to a worm gear, the lower end of the worm gear is meshed with a worm wheel, and one end of the rubber roller passes through the frame and is fixedly connected to the worm wheel;

[0018] Guide roller one and guide roller two are rotatably connected to the inner wall of the frame. Guide roller one is close to the right end of the frame, and guide roller two is close to the left end of the frame.

[0019] The present invention also provides a method for producing coated rubber rollers, applicable to the above-mentioned apparatus for producing coated rubber rollers based on wet paste preparation, comprising the following steps:

[0020] Step 1: Set the initial thickness and pressure parameters according to the viscosity of the polymer adhesive and the density of the carbon fiber cloth. Collect the pressure in real time through the pressure sensor and record the mapping relationship between the initial thickness and the corresponding pressure value simultaneously.

[0021] Step 2: Based on the images captured by the camera after scraping the adhesive, mark the abnormal pressure points and increase the sampling density to generate an image, coordinate, and pressure value mapping table to locate the problem area;

[0022] Step 3: Calculate the actual thickness of the adhesive layer based on the AI ​​segmentation model, distinguish between thicker and thinner areas by combining pressure data, establish a three-dimensional correlation model, and analyze the relationship between thickness, pressure, and viscosity.

[0023] Step 4: Based on the learning decision model and the thickness deviation and pressure suggestions, the optimal instructions are generated with the uniformity improvement rate as the core, and the target values ​​of servo motor displacement and scraper pressure are given.

[0024] Step 5: Drive the scraper blade to move using a servo motor, record and correct trajectory fluctuations in real time to ensure stable pressure and uniform scraping.

[0025] Step 6: Monitor the adjusted pressure using a pressure sensor. If the deviation is too large, the servo motor controls the scraper to retract. Record the stabilization time to ensure the pressure is within the target range.

[0026] Step 7: Establish a multi-dimensional database, optimize the model through machine learning, fine-tune the AI ​​segmentation model with new defect samples, and continuously improve the system's ability to identify and adjust complex defects.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] (1) This solution uses multiple sensors to monitor and adjust the glue scraping parameters in real time, uses AI models to accurately calculate the glue layer thickness deviation, and combines pressure and viscosity data to generate optimization suggestions, thereby improving the uniformity of plastic lamination, reducing quality problems such as glue layer cracking or insufficient adhesion, and enhancing the overall performance of the product.

[0029] (2) This solution introduces a learning decision model and a servo motor control system to realize the intelligent application of glue. The model generates the optimal instructions based on the thickness deviation and pressure suggestions, automatically adjusts the motor displacement and the pressure of the glue scraper, improves production efficiency, reduces reliance on manual labor, and ensures stable and reliable production.

[0030] (3) This solution establishes a multi-dimensional database and continuously optimizes the model through machine learning. The system can learn to adapt to new environments and material properties, fine-tune the AI ​​model with new defect samples, improve the ability to identify and adjust complex defects, and maintain technological leadership and production efficiency. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0032] Figure 1 This is an overall structural appearance view of the present invention;

[0033] Figure 2 This is a schematic diagram of the internal structure of the frame of the present invention;

[0034] Figure 3 This is a schematic diagram of the structure of the glue scraper in this invention;

[0035] Figure 4 This is a rear view of the overall structure of the present invention;

[0036] Figure 5 This is a schematic diagram of the worm gear structure of the present invention;

[0037] Figure 6 This is a cross-sectional view of the scraper blade of the present invention.

[0038] Explanation of the labels in the diagram:

[0039] 1. Frame; 2. Adhesive tank; 3. Adhesive roller; 4. Guide roller; 5. Guide roller one; 6. Guide roller two; 7. Drive motor; 8. Worm gear; 9. Worm; 10. Support plate one; 11. Camera; 12. Support plate two; 13. Servo motor; 14. Lead screw; 15. Fixing block; 16. Mounting plate; 17. Adhesive scraper; 18. Limiting plate; 19. Pressure sensor; 20. Elastic block. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Please see Figures 1 to 6A production roller coating device based on wet paste application includes a frame 1, an adhesive tank 2 mounted on the frame 1, an adhesive roller 3 disposed inside the adhesive tank 2, a drive motor 7 mounted on the outer side of the frame 1 for driving the adhesive roller 3 to rotate, a guide roller 4 rotatably connected to the inner wall of the frame 1 directly above the adhesive roller 3, a scraper blade 17 disposed inside the frame 1 above the guide roller 4 for scraping the polymer adhesive onto the surface of carbon fiber cloth, a servo motor 13 mounted on the frame 1 near the scraper blade 17 for controlling the movement of the scraper blade 17, a pressure sensor 19 mounted on the scraper blade 17 for detecting the pressure applied by the scraper blade 17 to the carbon fiber cloth, and a viscosity sensor installed inside the adhesive tank 2. The device includes a viscosity sensor to detect the viscosity of the polymer adhesive in the adhesive tank 2, and a camera 11 installed on the frame 1 to acquire image data of the carbon fiber cloth surface after the glue scraper 17 scrapes the adhesive. Based on the preliminary scraping image data and the viscosity data of the polymer adhesive acquired by the camera 11, combined with the contact pressure value of the glue scraper 17 monitored in real time by the pressure sensor 19, the AI ​​segmentation model is used to analyze the image to calculate the adhesive layer thickness and its deviation. The learning decision model generates pressure adjustment suggestions and safety constraint commands to optimize the uniformity of scraping based on the adhesive layer thickness deviation, the current contact pressure value, and the viscosity data of the polymer adhesive. It outputs the displacement target value of the servo motor 13 and the pressure target value of the glue scraper 17, and the servo motor 13 dynamically adjusts the movement of the glue scraper 17 accordingly.

[0042] A support plate 10 is fixedly installed at the lower end of the camera 11. Limiting plates 18 are provided at both the upper and lower ends of the glue scraper 17. The front and rear ends of the limiting plates 18 are fixedly connected to the inner wall of the frame 1. An installation plate 16 is provided on the left side of the glue scraper 17. A pressure sensor 19 is located between the glue scraper 17 and the installation plate 16. The pressure sensor 19 is fixedly connected to the installation plate 16. An elastic block 20 is fixedly connected between the glue scraper 17 and the installation plate 16. A fixing block 15 is fixedly installed on the left side of the installation plate 16. A lead screw 14 is threaded onto the fixing block 15. The left end of the lead screw 14 is fixedly connected to the output end of the servo motor 13. A support plate 12 is fixedly connected to the lower end of the servo motor 13. The front and rear ends of the support plate 10 and the support plate 12 are fixedly connected to the inner wall of the frame 1.

[0043] The drive motor 7 is fixedly installed on the outer side of the frame 1. The output shaft of the drive motor 7 is fixedly connected to the worm 9. The lower end of the worm 9 is meshed with the worm wheel 8. One end of the rubber roller 3 passes through the frame 1 and is fixedly connected to the worm wheel 8. Guide roller 1 5 and guide roller 2 6 are also rotatably connected to the inner wall of the frame 1. Guide roller 1 5 is close to the right side of the frame 1, and guide roller 2 6 is close to the left side of the frame 1.

[0044] In use, the carbon fiber cloth is passed over the guide roller 5, then between the rubber roller 3 and the guide roller 4, and then over the guide roller 6 before being guided onto the laminating machine for lamination. During operation, the carbon fiber cloth continuously passes between the rubber roller 3 and the guide roller 4. The worm gear 9 is driven to rotate by the drive motor 7. When the worm gear 9 rotates, it drives the worm wheel 8 to rotate. When the worm wheel 8 rotates, it drives the rubber roller 3 to rotate. Because the lower half of the rubber roller 3 is immersed in the adhesive tank 2 and in contact with the polymer adhesive in the adhesive tank 2, and the upper end of the rubber roller 3 is in contact with the carbon fiber cloth directly above, the rubber roller 3 will apply the polymer adhesive in the adhesive tank 2 to the carbon fiber cloth when it rotates. After being coated, the carbon fiber cloth will pass through the scraper 17 when it moves, and the scraper 17 will scrape the polymer adhesive on the surface of the carbon fiber cloth evenly. The evenly coated carbon fiber cloth continues to move, and then the adhesive layer of the carbon fiber cloth is photographed by camera 11 to obtain real-time image data of the adhesive layer of the carbon fiber cloth. Based on the image data, the thickness deviation value of each area and defect marks, such as air bubbles, are obtained. Then, based on the feedback information, the servo motor 13 is controlled to control the forward and reverse rotation of the lead screw 14. When the lead screw 14 rotates forward and reverse, it drives the fixing block 15 to move left and right along the lead screw 14. When the fixing block 15 moves left and right, it drives the mounting plate 16 to move synchronously. When the mounting plate 16 moves, it drives the squeegee 17 to move via the elastic block 20. When the squeegee 17 moves to the right, it contacts the carbon fiber cloth and applies a pressure to it. Simultaneously, the pressure sensor 19 detects this pressure. When the squeegee 17 moves to the left, it reduces or eliminates the pressure, and does not contact the cloth. By dynamically adjusting the position of the squeegee 17 based on the acquired image data, the uniformity of the polymer adhesive coating thickness on the carbon fiber cloth surface is ensured. Two sets of limiting plates 18 restrict the movement of the squeegee 17, ensuring that it can only move left or right along the restricted direction.

[0045] In some embodiments of the present invention, the linkage parameters between the initial thickness of the polymer adhesive on the carbon fiber cloth and the state of the polymer adhesive are set in advance, and the pressure monitoring threshold of the scraper blade 17 is set. The contact pressure is collected in real time by the pressure sensor 19, and the mapping relationship between the initial thickness and the corresponding pressure value is recorded simultaneously.

[0046] The camera continuously captures images of the adhesive layer after initial scraping using camera 11. Based on the detected pressure anomaly points, the sampling point density is increased at the corresponding positions in the image. The image pixels are then bound to the actual position and pressure value of the carbon fiber cloth using a spatiotemporal alignment algorithm, generating a mapping table of image pixels, actual coordinates, and pressure values.

[0047] By adopting the above technical solution, firstly, based on the viscosity of the polymer adhesive, the material density of the carbon fiber cloth (e.g., 100-300 g / m²), and the preset target adhesive layer thickness (e.g., 0.15-0.25 mm), a linkage parameter between the initial thickness of the polymer adhesive on the carbon fiber cloth and the state of the polymer adhesive is set. For example, when the viscosity of the polymer adhesive increases by 10%, the corresponding initial thickness needs to be reduced by 8% to avoid polymer adhesive accumulation. Based on this, combined with the tensile strength of the carbon fiber cloth, a pressure monitoring threshold for the squeegee 17 is set. For example, the minimum pressure is not less than 0.6 N to prevent incomplete squeegee application, resulting in uneven polymer adhesive coating thickness after squeegee application; the maximum pressure does not exceed 1.4 N to avoid damaging the carbon fiber cloth. These parameters are recorded in the control system as a benchmark for subsequent pressure monitoring and thickness adjustment. After the adhesive roller 3 applies the polymer adhesive to the surface of the carbon fiber cloth, the carbon fiber cloth, carrying the initial adhesive layer, enters the working area of ​​the scraper 17. At this time, the pressure sensor 19 begins to collect the contact pressure between the scraper 17 and the carbon fiber cloth in real time, setting an interval time, for example, a sampling interval of 0.1 seconds. The initial thickness data and pressure data at the same time and position are automatically bound together to generate a mapping relationship table between the initial thickness and the contact pressure. For example, data such as a thickness of 0.2 mm corresponding to a pressure of 0.9 N and a thickness of 0.25 mm corresponding to a pressure of 1.1 N are recorded. These mapping relationships will serve as a reference for judging whether the subsequent pressure is normal. For example, when the thickness is 0.2 mm but the pressure is lower than 0.8 N, it is determined to be an abnormal pressure.

[0048] Camera 11 starts capturing images after the carbon fiber cloth passes the adhesive scraper 17. The shooting frequency matches the speed of the carbon fiber cloth's movement; for example, when the carbon fiber cloth moves 0.1 meters per second, 2 frames are captured per second to ensure complete coverage of the adhesive layer surface. At this time, the generated initial thickness and contact pressure mapping table is called to filter out pressure anomalies, i.e., pressure values ​​exceeding the 0.6-1.4N threshold or deviations from the standard pressure for the corresponding thickness exceeding 0.2N. For these anomalies, camera 11 is automatically instructed to increase the sampling point density in the corresponding image area. For example, 5 pixels are collected per square centimeter in normal areas, while 15 pixels are collected in abnormal areas. Higher density sampling captures details of the adhesive layer at abnormal locations, such as bubbles or local over-thickness. The actual movement position of the carbon fiber cloth is determined by the positioning marks on the edge of the carbon fiber cloth, such as the preset black positioning line. Combined with the installation coordinates of camera 11, the actual physical coordinates of the carbon fiber cloth corresponding to each pixel in the image are calculated, such as the position of the upper left pixel on the carbon fiber cloth (0.2m, 0.05m). Subsequently, image data marked with anomalies and corresponding pressure data are retrieved, and the three are bound together using a spatiotemporal alignment algorithm: each image pixel not only corresponds to the actual coordinates of the carbon fiber cloth, but is also associated with the recorded contact pressure value at that coordinate point. Finally, a mapping table of image pixels, actual coordinates, and pressure values ​​is generated. Each entry in the table directly reflects the specific location on the carbon fiber cloth corresponding to the adhesive layer state displayed by a pixel, as well as the scraping pressure at that location, providing a direct basis for subsequent adhesive layer thickness judgment and adjustment of the scraper blade 17.

[0049] In some embodiments of the present invention, the image is segmented into adhesive regions according to the AI ​​segmentation model, and the actual thickness of each region is calculated by grayscale value analysis; the pressure data of the corresponding position is matched with the coordinate mapping table to distinguish the correlation between thicker and thinner thickness; at the same time, a three-dimensional correlation model of thickness deviation, scraping pressure and polymer adhesive viscosity is established based on the state data of polymer adhesive, and the pressure adjustment suggestion of scraper blade 17 is obtained based on the three-dimensional correlation model.

[0050] By adopting the above technical solution, the image is processed using an AI segmentation model. First, edge detection is used to identify the boundary between the adhesive area and the non-adhesive area of ​​the carbon fiber cloth, with segmentation accuracy controlled within ±1 pixel. Then, grayscale data within the adhesive area is extracted. The thicker the adhesive layer, the less light is reflected to camera 11 after passing through the adhesive layer, and the lower the grayscale value. For example, a thickness of 0.2mm corresponds to a grayscale value of 120, and 0.25mm corresponds to a grayscale value of 90. Combining the previously calibrated grayscale value and the correspondence between the actual thickness, which was obtained in advance through experiments with standard adhesive layer images of known thickness, the actual thickness value corresponding to each pixel in the adhesive area is calculated. This is then associated with an image pixel, actual coordinate, and pressure value mapping table to mark the thickness data at each actual coordinate position.

[0051] The pressure values ​​corresponding to each actual coordinate are extracted from the mapping table and bound one-to-one with the adhesive layer thickness data at each actual coordinate location. For example, coordinates (0.2m, 0.05m) correspond to a thickness of 0.22mm and a pressure of 0.8N. Through statistical analysis, areas with excessive thickness (exceeding the target thickness by more than 0.02mm) and areas with insufficient thickness (below the target thickness by more than 0.02mm) are identified. The pressure distribution characteristics of the two types of areas are statistically analyzed. For example, it was found that in areas with excessive thickness, 80% of the pressure values ​​are more than 0.15N lower than the standard pressure for the corresponding thickness, while in areas with insufficient thickness, 75% of the pressure values ​​are more than 0.1N higher than the standard pressure for the corresponding thickness, forming a clear correlation between thickness deviation and pressure.

[0052] The viscosity data of polymeric adhesives is linked with thickness and pressure data. For example, at the same time and coordinate, a thickness of 0.22 mm and a pressure of 0.8 N correspond to a viscosity of 300 mPa·s. A three-dimensional correlation model is then trained using historical data. The model analyzes the changes in the relationship between thickness and pressure under different viscosities. For instance, when the viscosity increases from 300 mPa·s to 330 mPa·s (an increase of 10%), the original rule of a thickness of 0.2 mm corresponding to a pressure of 0.9 N will be adjusted to a thickness of 0.2 mm corresponding to a pressure of 1.0 N, because high-viscosity polymeric adhesives require greater pressure to achieve the target thickness. The model determines the cause of the current thickness deviation by using a 3D model. For example, if the current thickness is 0.03mm too thick, the pressure is 0.7N, which is 0.2N lower than the standard value, and the viscosity is 350mPa·s, which is 300mPa·s higher than the benchmark value, the model will determine that it is necessary to simultaneously compensate for the insufficient pressure and deal with the influence of high viscosity. Combined with the adjustment coefficient in the model, for example, for every 10% increase in viscosity, the pressure needs to be increased by 8%, the target pressure is calculated. This calculated target pressure is the pressure adjustment suggestion for the scraper blade 17.

[0053] Regarding the construction of AI segmentation models:

[0054] A large number of image samples from real-world scenarios are collected in advance, covering different lighting conditions, polymer adhesive thickness from thin to thick, carbon fiber cloth and surface conditions, etc., to ensure sample diversity. Then, the boundaries between the adhesive area and the non-adhesive area are marked manually or with the help of semi-automatic tools, which generates mask labels. The adhesive area is marked with a specific identifier, and the non-adhesive area of ​​the carbon fiber cloth is marked with another identifier, ensuring that the labels correspond one-to-one with the pixels of the original image, and the boundary marking accuracy is as close to reality as possible.

[0055] Then, U-Net was chosen to build the model architecture. This type of architecture works through an encoder and decoder structure: the encoder part extracts deep features of the image through multi-layer convolution operations, such as the gray-level difference and texture changes between the glue area and the carbon fiber cloth; the decoder part fuses the shallow detail features and deep global features output by the encoder through skip connections, gradually restores the image size, and finally outputs the segmentation result of the same size as the original image, that is, the probability of each pixel belonging to the glue area or the non-glue area.

[0056] The labeled samples are divided into training and validation sets. During training, the model continuously adjusts its internal parameters by calculating the difference between the predicted segmentation result and the labeled value (i.e., the loss value), gradually bringing the prediction closer to the true boundary. The process focuses on the loss value in boundary regions, and parameters such as the number of network layers and convolutional kernel size are adjusted to enhance the model's ability to capture edge details. If inaccurate segmentation occurs in certain scenes, these samples are collected separately for targeted supplementary training to further optimize the model until it stably meets the actual segmentation requirements before deployment.

[0057] In some embodiments of the present invention, a learning decision model is established in advance. The learning decision model is used to receive thickness deviation data from image analysis and pressure adjustment suggestions for the scraper blade 17. The thickness uniformity improvement rate is used as the core reward function. The optimal instruction is generated in combination with safety constraints, and the target displacement value of the servo motor 13 and the target pressure value of the scraper blade 17 are given.

[0058] By adopting the above technical solution, the learning decision model first standardizes the input data, converting thickness deviation data into a severity level. For example, a deviation of 0.01-0.03mm is considered mild, 0.03-0.05mm is moderate, and a deviation greater than 0.05mm is severe. The actual coordinates of the severe deviation area are also marked. The pressure adjustment suggestion is broken down into a base pressure value and adjustment direction. For example, a suggestion to adjust from 0.8N to 1.1N is broken down into a base value of 0.8N and a positive adjustment of 0.3N. At the same time, historical data is correlated to extract the adjustment effect of similar deviations within the past 5 minutes. For example, in the past, after a moderate deviation was adjusted by 0.2N, the uniformity improvement rate was approximately 30%.

[0059] The expected benefit of the current adjustment plan is calculated using the thickness uniformity improvement rate as the reward function, specifically evaluated through three dimensions: First, the coverage of severely biased areas; if the adjustment can cover more than 80% of severely biased areas, a base reward of 30 points is added. Second, the matching degree between the adjustment range and historical results; for example, if the current recommended adjustment of 0.3N matches the optimal adjustment range of 0.25-0.35N for the same type of deviation in history, a reward of 20 points is added. Third, the predicted stability after adjustment; referring to the pressure fluctuation value after the same adjustment range in historical data, if the fluctuation is <0.05N, a reward of 15 points is added. The higher the total reward score, the greater the potential of the adjustment plan to improve uniformity. Regarding the calculation of the thickness uniformity improvement rate, for example, if the thickness fluctuation before optimization is 0.5mm and the fluctuation after optimization is 0.3mm, the improvement amount is 0.2mm, and the improvement rate is (0.2÷0.5)×100%=40%, meaning that the uniformity has been improved by 40%.

[0060] The learning decision-making model invokes preset safety constraint rules to verify the feasibility of the current adjustment plan. These safety constraints include two categories: first, physical limit constraints, such as the maximum pressure of the scraper blade 17 not exceeding 1.5N to prevent damage to the carbon fiber cloth; and the servo motor 13 driving the scraper blade 17 within a displacement range of ±0.02mm to avoid exceeding the mechanical stroke. Second, process stability constraints, such as a single pressure adjustment amplitude not exceeding 0.3N to prevent cracks in the adhesive layer due to sudden pressure changes; and a displacement adjustment speed not exceeding 0.01mm / ms to avoid mechanical vibration. If the suggested pressure adjustment exceeds the constraints, such as a suggested adjustment of 0.4N, a correction mechanism is automatically triggered. The adjustment amplitude is first reduced to 0.3N, and then the adjustment area is adjusted based on the reward score, prioritizing coverage of the most severely biased areas with the highest reward scores.

[0061] The learning decision-making model determines the adjustment priority based on the reward score details. If the coverage score of the severely deviated area is the highest, then that area is prioritized. Combined with the adjustment plan after safety constraints, the specific target value is calculated. For example, for the severely deviated area at coordinates (0.5m, 0.1m), based on the correlation that a thickness deviation of 0.04mm corresponds to a pressure increase of 0.2N, and combined with the maximum adjustment range of 0.3N under the safety constraint, the target pressure value for the scraper blade 17 in this area is determined to be 1.0N, with the current pressure at 0.8N plus an adjustment of 0.2N. Simultaneously, through the pressure and displacement correlation table, based on the previously calibrated rule that each 0.1N increase in pressure corresponds to a 0.008mm increase in displacement, the servo motor 13 needs to move the scraper blade 17 forward by 0.016mm (0.2N × 0.008mm / 0.1N), which is used as the displacement target value.

[0062] Compare the actual uniformity improvement rate with the expected reward score. For example, if the area of ​​the severely deviated region is reduced by 40% after actual adjustment, and the actual improvement rate is higher than expected (e.g., a reward score of 60 points corresponds to an actual improvement of 40%, which is higher than 30% of the historical average score), then record the parameters of the adjustment scheme, such as displacement of 0.016mm and pressure of 1.0N, as a high-quality case. If the actual improvement rate is lower than expected, then mark the cause of the deviation, such as the viscosity fluctuation of the polymer adhesive causing the adjustment effect to be discounted, and update the process stability constraint in the safety constraints, such as adding the adjustment range limit when the viscosity fluctuates.

[0063] Regarding the construction of learning decision models:

[0064] The model receives thickness deviation data from image analysis and suggestions for adjusting the pressure of the scraper blade beforehand. It categorizes thickness deviations into three levels: light, moderate, and severe (0.01-0.03mm, 0.03-0.05mm, and >0.05mm), and simultaneously marks the actual coordinates of the severe deviation area. The pressure adjustment suggestion is broken down into a base pressure value and adjustment direction; for example, adjusting from 0.8N to 1.1N is broken down into a base value of 0.8N and a positive adjustment of 0.3N. It also correlates with historical adjustment effects of similar deviations within the past 5 minutes; for example, data showing a 30% improvement in uniformity after a 0.2N adjustment for moderate deviations provides a reference for current decision-making. Based on this, a reward score is calculated from three dimensions using the thickness uniformity improvement rate as the core reward function: 30 points for covering more than 80% of the severe area, 20 points for matching the adjustment range to the historical best range, and 15 points for pressure fluctuation <0.05N after adjustment. The total score quantifies the improvement potential of the adjustment scheme, forming a decision-making benchmark.

[0065] Two types of safety constraints are preset: physical limit constraints specify that the pressure of the scraper blade should not exceed 1.5N and the displacement range should be ±0.02mm to prevent mechanical damage or exceeding the stroke; process stability constraints limit the single pressure adjustment to no more than 0.3N and the displacement speed to no more than 0.01mm / ms to avoid adhesive layer cracking or mechanical vibration. When the adjustment suggestion exceeds the constraints, such as a pressure adjustment reaching 0.4N, the model automatically triggers correction: first, the amplitude is compressed to the safe range of 0.3N, and then the area of ​​action is adjusted according to the priority of reward scores, prioritizing coverage of the most severely biased areas with the highest rewards, ensuring that the decision optimizes the target within the safe framework, and achieving a balance between safety and adjustment effect.

[0066] Adjustment priorities are determined based on the detailed reward scores. For example, if the coverage score for a severely biased area is the highest, that area is prioritized. The specific target value is calculated by combining the correlation between thickness deviation and pressure, as well as the correlation table between pressure and displacement. After adjustment, the actual uniformity improvement rate is compared with the expected reward score. Solutions that exceed expectations are recorded as high-quality cases, while those that do not meet expectations are analyzed for the reasons, and the stability constraints of the process are updated. Through a closed loop of data input, decision output, effect feedback, and parameter iteration, the model's decision accuracy and environmental adaptability are continuously improved.

[0067] In some embodiments of the present invention, the scraper blade 17 is driven to move according to the displacement target value; the movement trajectory of the scraper blade 17 is recorded synchronously; if the trajectory fluctuates, feedback is provided in real time and the control of the servo motor 13 is smoothly corrected.

[0068] By adopting the above technical solution, the displacement target value is first decomposed into the direction of movement, the total displacement, and the duration of action. For example, the movement is initiated when the 0.5m-0.6m section of carbon fiber cloth enters the adhesive scraping area. Combined with the pressure target value, a preset displacement and pressure linkage program is used. For example, the movement is initiated at a low speed of 0.002mm / ms in the initial stage, while the pressure sensor 19 monitors the contact pressure in real time to ensure that the pressure is not lower than 0.9N during the movement. The servo motor 13 drives the lead screw 14 to rotate, which in turn drives the adhesive scraper 17 to move along the target trajectory. During the movement, the real-time position is recorded every 0.001mm to ensure positional accuracy.

[0069] The location data is integrated chronologically to generate a complete movement trajectory curve, with the horizontal axis representing time and the vertical axis representing position. Key features are extracted using trajectory analysis tools: First, the deviation between the actual displacement and the target displacement, such as a target movement of 0.016mm and an actual movement of 0.0158mm, resulting in a deviation of 0.0002mm; second, the trajectory fluctuation amplitude, for example, calculating the position fluctuation value within each 0.001mm movement segment, marking a fluctuation point if the fluctuation exceeds 0.0003mm; third, the correlation between fluctuation and pressure, matching the timestamps corresponding to fluctuation points with pressure data to determine whether the trajectory fluctuation was caused by a sudden change in pressure, while also recording the trajectory completion time and comparing it with the preset standard time.

[0070] The system invokes preset trajectory fluctuation judgment criteria: first, a single fluctuation threshold, such as a fluctuation amplitude > 0.0005mm, is considered abnormal fluctuation; second, the number of consecutive fluctuations, such as more than 3 fluctuation points within 1 second, is judged as continuous abnormality. Fluctuation points are classified: if the fluctuation is related to pressure, such as a sudden drop in pressure accompanied by trajectory fluctuation, it is marked as pressure-related fluctuation; if the pressure is stable but the trajectory fluctuates, such as due to mechanical friction of the lead screw 14, it is marked as mechanical fluctuation. Simultaneously, considering the thickness requirements of the affected area, such as the 0.5m-0.6m section being a region of severe deviation with lower fluctuation tolerance, the judgment strictness is adjusted, reducing the regional fluctuation threshold to 0.0003mm.

[0071] If the trajectory is determined to be normal with no abnormal fluctuations, servo motor 13 maintains the current control parameters. If abnormal fluctuations occur, a correction mechanism is activated: For pressure-related fluctuations, such as those caused by a sudden drop in pressure, the moving speed is finely adjusted using pressure and displacement compensation algorithms. For example, if the pressure is below 0.9N, the moving speed is temporarily reduced to 0.001mm / ms, and restored after the pressure recovers. For mechanical fluctuations, such as friction of the lead screw 14, a lubrication compensation signal for the lead screw 14 is automatically added, and a small reverse pulse is output by servo motor 13 to counteract frictional resistance. During the correction process, the magnitude of each correction is strictly controlled, such as each adjustment not exceeding 0.0005mm, and the trajectory changes after correction are recorded simultaneously.

[0072] The corrected trajectory fluctuation amplitude is compared with that before correction. For example, after lubrication compensation, the fluctuation amplitude of mechanical fluctuations decreases from 0.0006mm to 0.0002mm to evaluate the effectiveness of the correction. Simultaneously, thickness uniformity data is correlated; if the thickness deviation in this area decreases from 0.04mm to 0.015mm, it indicates that the correction directly contributes to improving uniformity. Effective correction parameters, such as the speed adjustment value for pressure-related fluctuations, are recorded in the trajectory amendment example library as a priority correction scheme for subsequent similar fluctuations.

[0073] In some embodiments of the present invention, the pressure sensor 19 monitors the adjusted pressure, calculates the deviation between the pressure and the target pressure value, and compares the deviation with a preset deviation judgment threshold. If the absolute value of the deviation is greater than the deviation judgment threshold, it is fed back to the servo motor 13 in real time, and the servo motor 13 controls the scraper blade 17 to move backward, while the pressure stabilization time is recorded.

[0074] By adopting the above technical solution, the pressure sensor 19 continuously monitors the contact pressure between the scraper blade 17 and the carbon fiber cloth at a preset sampling frequency, such as generating a pressure data point every 0.001 seconds. The raw pressure data is preprocessed to filter out burr data caused by instantaneous vibration, such as removing outliers with a deviation exceeding 0.01N from three adjacent data points. Simultaneously, the data is linked to the current position of the scraper blade 17 according to a timestamp; for example, at 10:00:00, a pressure of 1.02N corresponds to a position of 0.016mm on the scraper blade 17. The actual current pressure value is extracted from the pressure, position, and time-related data table and compared point-by-point with the target pressure value to calculate the deviation value: deviation value = actual pressure - target pressure. Set a deviation judgment threshold: when the absolute value of the deviation is ≤0.02N, the pressure is judged to be normal; when the deviation value is >0.02N, the pressure is too high or the deviation value is <-0.02N, the pressure is too low, the deviation is judged to be greater than the target value, and the duration of the deviation is marked. If the deviation of 3 consecutive data points exceeds the threshold, it is judged to be a continuous deviation.

[0075] If the pressure is determined to be normal, the current position of the scraper blade 17 is maintained, and no adjustment is triggered. If the deviation is determined to be greater than the target value, a corresponding instruction is generated according to the deviation type: when the pressure is too high, the deviation value > 0.02N, the required retreat distance is calculated, referring to the pressure and distance association database. For example, for every 0.01N increase in pressure, a retreat of 0.0008mm is required. For instance, when the pressure is 1.05N, a deviation of 0.05N requires a retreat of 0.004mm. Simultaneously, the retreat speed is set based on the duration of the deviation. If the continuous deviation exceeds 0.003 seconds, a low-speed retreat of 0.001mm / ms is used to avoid a sudden drop in pressure. After receiving the retreat instruction, the servo motor 13 drives the lead screw 14 to rotate in the opposite direction at the set speed, causing the scraper blade 17 to gradually retreat. During the retreat process, the pressure sensor 19 synchronously monitors the pressure change, recording the pressure value every 0.0001mm of retreat. For example, if the pressure drops to 1.03N when retreating 0.001mm, and drops to 1.01N when retreating 0.002mm. The system compares the deviation between the current pressure and the target value in real time. When the deviation is reduced to within 0.01N, it approaches the target value and automatically reduces the retreat speed to 0.0005mm / ms for fine-tuning.

[0076] Once the pressure deviation stabilizes within ±0.01N, reaching the target pressure fluctuation range, a stabilization judgment timer is initiated: Five consecutive sampling cycles are monitored, totaling 0.005 seconds. If the pressure does not exceed ±0.01N, the pressure is considered stable. The total time from the issuance of the retraction command to pressure stabilization is recorded; for example, from 10:00:00.003 to 10:00:00.010, the stabilization time is 0.007 seconds. Simultaneously, the position of the scraper blade 17 at the point of stabilization is recorded; for example, after retraction of 0.003mm, the position is 0.013mm. This position is compared with the target position to calculate the positioning accuracy.

[0077] Complete data from this adjustment is archived to the database, including initial pressure deviation, retreat distance, stabilization time, and final pressure. Data is stored categorized by deviation type and region, such as the case of excessive pressure adjustment in the 0.5m section. By analyzing the stabilization time differences in similar cases, optimization parameters are extracted: for example, if it is found that excessive pressure occurs repeatedly in a certain region and the stabilization time exceeds 0.01 seconds, the pressure and distance correlation database is updated, and the retreat distance coefficient for that region is appropriately increased, such as from 0.0008mm / 0.01N to 0.0009mm / 0.01N.

[0078] In some embodiments of the present invention, a database of pressure, images, adjustment trajectories, and thickness is established, an optimization model is trained through machine learning, and the AI ​​segmentation model is fine-tuned using new defect samples.

[0079] By adopting the above technical solution, pressure data is classified according to pressure value, deviation type, and stabilization time, such as 1.05N - too high - 0.007 seconds; image data is labeled according to defect type, defect area, and grayscale value distribution, such as too thick - 0.5cm², grayscale value 80-100; adjustment trajectory data is converted into features of movement distance, speed, and pressure change rate, such as 0.004mm, 0.001mm / ms, 0.02N / ms; thickness data is quantified according to actual thickness, target thickness, and deviation rate, such as 0.22mm, 0.20mm, 10%. A correlation is established through timestamps and carbon fiber cloth coordinates. For example, pressure, image, trajectory, and thickness data at the same coordinate (0.5m, 0.1m) at 10:00:00.000 are bound and stored in a structured database, with each type of data corresponding to a unique association ID.

[0080] Key features are extracted from the database: pressure fluctuation frequency and stable time distribution are extracted from pressure data; edge features of defect areas and grayscale gradient changes are extracted from image data; the correlation between adjustment amplitude and pressure response delay, and trajectory fluctuation and thickness deviation are extracted from trajectory data; and clustering features of deviation areas and thickness uniformity improvement rate are extracted from thickness data. Training samples are constructed based on input features and output results: pressure features, image features, and trajectory features are used as inputs, and thickness uniformity improvement rate is used as the output label. For example, an input pressure fluctuation of 0.02N, a thicker defect, and a trajectory fluctuation of 0.0005mm corresponds to an output improvement rate of 30%. Sample quality is also labeled; samples containing all four types of data are marked as high-quality samples.

[0081] A machine learning algorithm combining classification and regression is employed: first, a classification algorithm is used to learn the correspondence between defect types and optimal adjustment trajectories, such as a thickness defect corresponding to a retreat distance of 0.003-0.005mm; then, a regression algorithm is used to fit the pressure, trajectory, and thickness improvement rate function, such as predicting the thickness improvement effect under a certain combination of pressure and trajectory parameters. During training, weights are allocated according to the principle of prioritizing high-quality samples, with high-quality samples having twice the weight of ordinary samples. The model accuracy is validated every 1000 training samples, with a deviation of ≤5% between the predicted and actual improvement rates considered acceptable. If the accuracy does not meet the standard, high-weight features are strengthened through a feature importance ranking table, such as increasing the training proportion of samples related to pressure stabilization time.

[0082] New defect samples are manually annotated to determine defect types, boundaries, and features, such as mixed defects, bubbles in thicker areas, and abrupt grayscale changes. These are then added to the image database and associated with corresponding pressure, trajectory, and thickness data to form new defect, adjustment, and effect samples. The new samples are split into incremental training and test sets at a 1:4 ratio. The incremental training set is used to supplement the training of the optimized model, retaining the original model parameters and updating only the weights related to the new defects. The test set is used to verify the accuracy of new defect recognition, requiring ≥90%.

[0083] Novel edge features and defect texture features are extracted from new defect sample images, such as the gray-scale abrupt boundary of mixed defects and irregular light spots in bubble regions, as fine-tuning inputs. The convolutional layer weights of the original AI segmentation model are adjusted to enhance its sensitivity to new features. Mini-batch training is performed using new defect samples, with 50-100 images trained each time. After training, the effect is verified by the defect segmentation accuracy, such as the degree of overlap between the segmentation boundary and the actual defect boundary. If the segmentation accuracy of a certain type of new defect is <85%, the number of training times for that type of sample is increased by 20 images each time until the target is reached.

[0084] The fine-tuned segmentation model is used to identify defects in images of the new batch of carbon fiber fabric, outputting defect types and locations. The results are then input into an optimization model to predict the optimal adjustment trajectory and pressure parameters. After adjustment, the actual thickness data is collected and compared with the thickness increase rate predicted by the model to calculate the overall error, which should be ≤3%. If the error exceeds the standard, the database is traced back to find the cause of the deviation, such as changes in the adjustment effect due to the viscosity of the new polymer adhesive. Corresponding samples are then added to re-fine-tune the model.

[0085] The present invention also provides a method for producing coated rubber rollers, applicable to the above-mentioned apparatus for producing coated rubber rollers based on wet paste preparation, comprising the following steps:

[0086] Step 1: Set the initial thickness and pressure parameters according to the viscosity of the polymer adhesive and the density of the carbon fiber cloth. Collect the pressure in real time through the pressure sensor 19 and record the mapping relationship between the initial thickness and the corresponding pressure value simultaneously.

[0087] Step 2: Based on the image captured by camera 11 after scraping the adhesive, mark the abnormal pressure points and increase the sampling density to generate an image, coordinate and pressure value mapping table to locate the problem area;

[0088] Step 3: Calculate the actual thickness of the adhesive layer based on the AI ​​segmentation model, distinguish between thicker and thinner areas by combining pressure data, establish a three-dimensional correlation model, and analyze the relationship between thickness, pressure, and viscosity.

[0089] Step 4: Based on the learning decision model and the thickness deviation and pressure suggestions, the optimal instructions are generated with the uniformity improvement rate as the core, and the target values ​​of the displacement of the servo motor 13 and the pressure of the scraper blade 17 are given.

[0090] Step 5: Drive the scraper blade 17 to move via the servo motor 13, record and correct trajectory fluctuations in real time, and ensure stable pressure and uniform scraping.

[0091] Step 6: Monitor the adjusted pressure using pressure sensor 19. If the deviation is too large, the servo motor 13 controls the scraper blade 17 to retreat. Record the stabilization time to ensure that the pressure is within the target range.

[0092] Step 7: Establish a multi-dimensional database, optimize the model through machine learning, fine-tune the AI ​​segmentation model with new defect samples, and continuously improve the system's ability to identify and adjust complex defects.

[0093] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A production roller coating device based on wet paste forming, comprising a frame (1), an adhesive tank (2) installed on the frame (1), an adhesive roller (3) disposed inside the adhesive tank (2), a drive motor (7) installed on the outer side of the frame (1), the drive motor (7) being used to drive the adhesive roller (3) to rotate, and a guide roller (4) rotatably connected to the inner wall of the frame (1) directly above the adhesive roller (3). Its features are: A scraper blade (17) is installed above the guide roller (4) inside the frame (1). The scraper blade (17) is used to scrape the polymer adhesive on the surface of the carbon fiber cloth. A servo motor (13) is also installed on the frame (1) near the scraper blade (17). The servo motor (13) is used to control the movement of the scraper blade (17). A pressure sensor (19) is installed on the scraper blade (17). The pressure sensor (19) is used to detect the pressure applied by the scraper blade (17) to the carbon fiber cloth. A viscosity sensor is installed in the adhesive tank (2). The viscosity sensor is used to detect the viscosity of the polymer adhesive in the adhesive tank (2). A camera (1) is also installed on the frame (1). 1) The camera (11) is used to acquire image data of the carbon fiber cloth surface after the scraper (17) scrapes the glue. The preliminary scraping image data and the viscosity data of the polymer adhesive acquired by the camera (11) are combined with the contact pressure value of the scraper (17) monitored in real time by the pressure sensor (19). The AI ​​segmentation model is used to analyze the image to calculate the glue layer thickness and its deviation. The learning decision model generates optimized scraping pressure adjustment suggestions and safety constraint instructions based on the glue layer thickness deviation, the current contact pressure value and the viscosity data of the polymer adhesive. The displacement target value of the servo motor (13) and the pressure target value of the scraper (17) are output. The servo motor (13) dynamically adjusts the movement of the scraper (17) accordingly.

2. The production roller coating device based on wet paste preparation according to claim 1, characterized in that: The linkage parameters between the initial thickness of the polymer adhesive on the carbon fiber cloth and the state of the polymer adhesive are set in advance, and the pressure monitoring threshold of the scraper (17) is set. The contact pressure is collected in real time by the pressure sensor (19), and the mapping relationship between the initial thickness and the corresponding pressure value is recorded synchronously. The adhesive layer after initial scraping is continuously photographed by camera (11), and the sampling point density is increased at the corresponding position of the image based on the detected pressure anomaly point marking; The spatiotemporal alignment algorithm binds image pixels to the actual position and pressure value of carbon fiber cloth, generating a mapping table of image pixels, actual coordinates, and pressure values.

3. The production roller coating device based on wet paste preparation according to claim 2, characterized in that: The image is segmented into adhesive areas according to the AI ​​segmentation model, and the actual thickness of each area is calculated by gray value analysis. The pressure data of the corresponding position is matched with the coordinate mapping table to distinguish the correlation between thicker and thinner thickness. At the same time, a three-dimensional correlation model of thickness deviation, scraping pressure and viscosity of polymer adhesive is established based on the state data of polymer adhesive. The pressure adjustment suggestion of scraping knife (17) is derived based on the three-dimensional correlation model.

4. The production roller coating device based on wet paste preparation according to claim 3, characterized in that: A learning decision model is established to receive thickness deviation data from image analysis and pressure adjustment suggestions from the scraper (17). The thickness uniformity improvement rate is used as the core reward function. The optimal instruction is generated in combination with safety constraints, and the displacement target value of the servo motor (13) and the pressure target value of the scraper (17) are given.

5. The production roller coating device based on wet paste preparation according to claim 4, characterized in that: The scraper (17) is driven to move according to the displacement target value; the movement trajectory of the scraper (17) is recorded synchronously; if the trajectory fluctuates, feedback is given in real time and the control of the servo motor (13) is smoothly corrected.

6. The production roller coating device based on wet paste preparation according to claim 5, characterized in that: According to the pressure sensor (19) monitoring the adjusted pressure, the deviation value between the pressure and the pressure target value is calculated, and the deviation value is compared with the preset deviation judgment threshold. If the absolute value of the deviation value is greater than the deviation judgment threshold, it is fed back to the servo motor (13) in real time, and the scraper (17) is controlled to move backward by the servo motor (13), while the pressure stabilization time is recorded.

7. The production roller coating device based on wet paste preparation according to claim 6, characterized in that: Establish a database of pressure, images, adjustment trajectories, and thickness; train and optimize the model through machine learning; and fine-tune the AI ​​segmentation model using new defect samples.

8. The production roller coating device based on wet paste preparation according to claim 1, characterized in that: A support plate (10) is fixedly installed at the lower end of the camera (11). Limiting plates (18) are provided at both the upper and lower ends of the glue scraper (17). The front and rear ends of the limiting plates (18) are fixedly connected to the inner wall of the frame (1). A mounting plate (16) is provided on the left side of the glue scraper (17). A pressure sensor (19) is located between the glue scraper (17) and the mounting plate (16). The pressure sensor (19) is fixedly connected to the mounting plate (16). An elastic block (20) is fixedly connected between the mounting plate (16) and the mounting plate (16). A fixing block (15) is fixedly installed on the left side of the mounting plate (16). A screw rod (14) is threaded onto the fixing block (15). The left end of the screw rod (14) is fixedly connected to the output end of the servo motor (13). A support plate (12) is fixedly connected to the lower end of the servo motor (13). The front and rear ends of the support plate (10) and the support plate (12) are fixedly connected to the inner wall of the frame (1).

9. A film coating device for production rollers based on wet paste preparation according to claim 8, characterized in that: The drive motor (7) is fixedly installed on the outer side of the frame (1). The output shaft of the drive motor (7) is fixedly connected to a worm (9). The lower end of the worm (9) is meshed with a worm wheel (8). One end of the rubber roller (3) passes through the frame (1) and is fixedly connected to the worm wheel (8). The inner wall of the frame (1) is also rotatably connected to guide roller one (5) and guide roller two (6). Guide roller one (5) is close to the right end of the frame (1), and guide roller two (6) is close to the left end of the frame (1).

10. A method for producing coated rubber rollers, said method being performed by the rubber roller coating apparatus based on wet paste preparation as described in claim 9, characterized in that: Includes the following steps: Step 1: Set the initial thickness and pressure parameters according to the viscosity of the polymer adhesive and the density of the carbon fiber cloth. Collect the pressure in real time through the pressure sensor (19) and record the mapping relationship between the initial thickness and the corresponding pressure value simultaneously. Step 2: Based on the image captured by the camera (11) after scraping the glue, mark the pressure abnormal points and increase the sampling density to generate an image, coordinate and pressure value mapping table to locate the problem area; Step 3: Calculate the actual thickness of the adhesive layer based on the AI ​​segmentation model, distinguish between thicker and thinner areas by combining pressure data, establish a three-dimensional correlation model, and analyze the relationship between thickness, pressure, and viscosity. Step 4: Based on the learning decision model and the thickness deviation and pressure suggestions, the optimal instructions are generated with the uniformity improvement rate, and the target values ​​of the displacement of the servo motor (13) and the pressure of the scraper (17) are given. Step 5: Drive the scraper blade (17) to move via the servo motor (13), and record and correct trajectory fluctuations in real time; Step 6: Monitor the adjusted pressure using the pressure sensor (19). If the deviation is too large, the servo motor (13) controls the scraper (17) to move backward and records the stabilization time. Step 7: Establish a multi-dimensional database, optimize the model through machine learning, and fine-tune the AI ​​segmentation model with new defect samples.

Images