A method for identifying and correcting wrinkles of a roll-based package based on visual feedback
By combining visual feedback with reparameterized convolution and multimodal deep learning models, wrinkle defects in roll packaging can be identified and corrected in real time. This solves the problems of poor robustness of traditional detection and time-consuming deep learning computation, and achieves efficient and interpretable automated control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG SHILING INTELLIGENT TECHNOLOGY CO LTD
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-14
AI Technical Summary
Existing roll packaging equipment is prone to defects such as film wrinkles and bubbles during high-speed packaging. Relying on manual visual inspection results in a high rate of missed detection and slow feedback. Traditional machine vision inspection has poor robustness, and deep learning models are time-consuming to compute on resource-constrained equipment, making it difficult to meet real-time response requirements.
A visual feedback-based approach is adopted, combining reparameterized convolution and multimodal deep learning models to acquire images and process parameters in real time. The cross-attention attribution fusion model is used to identify wrinkle types and calculate correction increments to achieve automated correction.
It enables real-time wrinkle recognition and correction in high-speed production environments, improving detection reliability and production efficiency, reducing material waste, and giving the deep learning model interpretability.
Smart Images
Figure CN122391101A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing technology for roll packaging equipment, specifically a method for identifying and correcting wrinkle defects in roll packaging based on visual feedback. Background Technology
[0002] In industrial packaging processes for roll materials such as metal strip rolls, cable rolls, and textile rolls, stretch film is typically used for wrapping. Existing automated packaging equipment can achieve efficient mechanical operations. However, during high-speed packaging, defects such as film wrinkles and bubbles are easily generated due to uneven film tension, uneven roll surface, centering misalignment, or equipment vibration. Currently, these defects mainly rely on manual visual inspection, which suffers from high missed inspection rates, delayed feedback, and inability to adjust process parameters in real time. This leads to unstable packaging quality, material waste, and may affect subsequent palletizing and transportation.
[0003] With the development of intelligent manufacturing, introducing machine vision and real-time control into the packaging process has become a trend. Currently, some automated production lines have introduced automatic inspection based on traditional machine vision, using fixed-position industrial cameras combined with simple image processing algorithms based on threshold, edge detection, or template matching. An alarm is triggered when image features exceed the threshold. This method may be effective in simple scenarios with static or low speed and constant lighting, but it faces the following problems in the dynamic scenario of roll packaging: poor robustness, weak adaptability to changes in lighting, surface reflection of rolls, and different material textures, resulting in a high false alarm rate; it can only alarm, not diagnose, the system can only indicate that there is a defect, but cannot determine whether it is due to uneven tension, inaccurate correction, or other reasons, and maintenance and debugging still require a lot of manual troubleshooting. On the other hand, single-modal vision inspection based on deep learning uses convolutional neural networks or visual Transformers to classify and locate defects in images. Although these methods are superior to traditional algorithms in terms of accuracy, their direct application in this scenario still has bottlenecks, namely, high-precision models (such as Faster R-CNN, Swin Transformer) have a large number of parameters and are computationally time-consuming, making it difficult to meet the real-time response requirements on resource-constrained industrial edge devices.
[0004] Therefore, it is necessary to design a visual feedback-based method for identifying and correcting wrinkle defects in roll packaging that is both highly robust and real-time. Summary of the Invention
[0005] The purpose of this invention is to provide a method for identifying and correcting wrinkle defects in roll packaging based on visual feedback, so as to solve the problems mentioned in the background art.
[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for identifying and correcting wrinkle defects in roll packaging based on visual feedback, comprising the following steps: Step S100: During the roll wrapping process, simultaneously acquire real-time images of the film surface. And process feature parameters, using a unified timestamp to spatiotemporally align the surface image with the process feature parameters; Step S200: Run a lightweight perception algorithm based on reparameterized convolution to perform sliding window scanning on the surface image. When the gradient change rate of the detected region exceeds the dynamic threshold, it is determined to be a suspected defect, triggering a high-frequency sampling mechanism to obtain the image of the center region of the suspected defect and the sequence of process parameter changes within a set time window before and after the defect occurs. Step S300: Input the image of the suspected defect center region and the sequence of process parameter changes within a set time window T seconds before and after the defect occurrence into the multimodal deep learning model to obtain image feature branch vectors respectively. and process timing feature branch vector The model uses a cross-attention attribution fusion model to output the wrinkle type C, mask M, and root cause attribution score A.
[0007] Step S400: Based on the wrinkle type C, mask M, and root cause score A, calculate the quantitative correction increment value for the tension regulating valve or the correction actuator under the current defect state.
[0008] S500: The correction position adjustment amount is sent to the corresponding actuator for execution.
[0009] According to the above technical solution, step S100 further includes: Step S101: During the roll wrapping process, an industrial camera captures real-time images of the film surface. The programmable logic controller (PLC) acquires process characteristic parameters in real time, and the same pulse signal simultaneously triggers camera exposure and PLC acquisition. The process characteristic parameters include the tension value detected by the tension sensor. Main line speed and displacement of the correction mechanism ; Step S102: Based on image frames The precise exposure time is Find the data in the PLC that matches Two adjacent time points and After time alignment, the tension value corresponding to that frame is... The calculation formula is: ; Step S103: Perform spatial correction based on the physical distance D between the industrial camera and the tension sensor, then the tension value after frame alignment is obtained. In the formula To compensate for the time difference in the travel time of the thin film from the tension sensor to the camera position, .
[0010] According to the above technical solution, step S300 further includes: Step S301: Input the image of the suspected defect center region into the adaptive high-low frequency fusion AHLF module, wherein the high-frequency branch uses a local window attention model to capture wrinkle edge details, and the low-frequency branch uses global pooling to model the overall deformation of the thin film to obtain the image feature branch vector. ; Step S302: Set the tension value within a time window before and after the defect occurs. Main line speed and displacement of the correction mechanism The sequence is input into a one-dimensional temporal convolutional network to extract the process time sequence feature branch vector. ; Step S303: Convert the image feature branch vector and process timing feature branch vector The input is fed into the cross-attention attribution fusion model to obtain the wrinkle type C, segmentation mask M, and root cause attribution score A.
[0011] According to the above technical solution, the cross-attention attribution fusion model Image feature branch vector For the query term, the process time sequence feature branch vector The key-value items are fused to calculate the contribution of process fluctuations to wrinkles at specific locations. The relationship is expressed as follows: in For feature dimensions.
[0012] According to the above technical solution, step S303 further includes: Step S3031: Through the cross-attention attribution fusion model A multimodal fusion feature matrix is obtained. The multimodal fusion feature matrix After performing global average pooling, the input is a linear classification layer. By calculating the probability distribution of each category, the defect category C is output. Step S3032: Use a deconvolutional network to process the multimodal fusion feature matrix. Perform spatial dimension restoration, generate a binarized mask M of the same size as the original image, and output a pixel-level defect segmentation map; Step S3033: Extract the cross-attention attribution fusion model The weight coefficient matrix is calculated, and the root cause attribution score A is output by statistically analyzing the contribution rate of each process parameter feature channel to the salient region of the image.
[0013] According to the above technical solution, step S400 further includes: Step S401: Calculate the severity index S of the current defect based on the wrinkle type C and the mask M; Step S402: Calculate the control increment under multimodal guidance by combining the root cause attribution score A. The calculation formula is: Where i represents the actuator, i=1 is the tension regulating valve, i=2 is the correction actuator; It locks the sensitivity gain of each of the actuators. It is a direction determination function.
[0014] According to the above technical solution, the system includes the following: The system includes a perception module, a high-speed recognition module, a precise diagnosis module, a decision control module, and a correction execution module, wherein... The sensing module, including a camera, a light source, and a sensor gateway, is configured to simultaneously acquire thin film surface images and real-time process parameters, and perform spatiotemporal alignment processing on the data. The high-speed recognition module is communicatively connected to the perception module and is configured to run a lightweight, heavily parameterized network to monitor the image stream in real time and trigger a warning of suspected defects. The precise diagnosis module is communicatively connected to the high-speed identification module and is configured to deploy a multimodal deep learning model for classifying, segmenting, and performing root cause analysis based on cross-attention for suspected defects. The decision control module is communicatively connected to the precise diagnosis module and is configured to calculate and generate quantitative adjustment instructions for each actuator based on the diagnosis results. The correction execution module, together with the decision control module, including the tension adjustment mechanism and / or the correction execution mechanism, is configured to receive and execute the quantitative adjustment command.
[0015] According to the above technical solution, the light source of the sensing module adopts a low-angle grazing strip light source array to enhance the contrast of the wrinkle features on the thin film surface.
[0016] The present invention also includes an electronic device, comprising a processor and a memory, wherein the processor runs a computer program or code stored in the memory to implement the above-described method for identifying and correcting wrinkles in roll packaging based on visual feedback.
[0017] The present invention also includes a computer-readable storage medium for storing a computer program or code, characterized in that, when the computer program or code is executed by a processor, it implements the above-described method for identifying and correcting wrinkles in roll-type packaging based on visual feedback. Compared with the prior art, the beneficial effects achieved by the present invention are: This invention compresses a complex multi-path training network into a highly efficient single-path matrix inference network through reparameterization technology, significantly improving inference speed and reducing memory usage without sacrificing accuracy, ensuring real-time response for wrinkle recognition in high-speed production environments. It utilizes a linear velocity adaptive spatiotemporal feature association mechanism to accurately align visual images with equipment status time-series data, and employs an AHLF module to extract high- and low-frequency features to force recognition results to conform to physical laws, effectively shielding against purely optical interference and solving recognition difficulties caused by the transparency, reflectivity, and multi-scale nature of packaging films, significantly enhancing the system's detection reliability. Simultaneously, by analyzing the cross-attention attribution fusion model, abstract features are transformed into interpretable root cause attribution scores, endowing the deep learning model with interpretability, enabling precise matching of adjustment values to defect causes, ultimately achieving a complete closed loop from recognition to wrinkle elimination, greatly improving control accuracy and the intelligent production efficiency of the production line. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system module composition of the present invention; Figure 3 This is a schematic diagram of step S300 of the present invention. Detailed Implementation
[0019] 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.
[0020] Please see Figure 1 This invention provides a technical solution: a method for identifying and correcting wrinkle defects in rolled packaging based on visual feedback, comprising the following steps: Step S100: During the roll wrapping process, simultaneously acquire real-time images of the film surface. The surface image and process feature parameters are spatiotemporally aligned using a unified timestamp; this step aims to solve the data misalignment problem caused by inconsistent sensor sampling frequencies and differences in physical installation locations.
[0021] Further, step S100 includes: Step S101 Data Acquisition: On the thin film winding line, image frames are acquired using an imaging system consisting of a low-angle grazing strip light source array. A rotary encoder synchronously generates pulse signals to trigger the PLC to capture the process parameter sequence in real time. The minimum set of process parameters includes: the process characteristic parameters include the tension value detected by the tension sensor. Main line speed and displacement of the correction mechanism .
[0022] Step S102 Timestamp Alignment: Based on image frames The precise exposure time is Find the data in the PLC that matches Two adjacent time points and After time alignment, the tension value corresponding to that frame is... The calculation formula is: Since the PLC sampling frequency (e.g., 500Hz) is much higher than the camera frame rate (e.g., 60Hz), linear interpolation alignment is performed using the above formula to resolve the time mismatch between the PLC's high-frequency data and the camera's low-frequency data. Each frame is assigned a precise physical parameter "snapshot," eliminating analysis errors caused by the sampling frequency difference.
[0023] Step S103 Physical Space Compensation: Spatial correction is performed based on the physical distance D between the industrial camera and the tension sensor, then the tension value after frame alignment is obtained. In the formula To compensate for the time difference in the travel time of the thin film from the tension sensor to the camera position, .
[0024] There is a physical distance D between the camera mounting point and the sensor mounting point. The tension currently sensed by the sensor actually acts on the thin film outside of this distance D. By compensating for the time difference in the thin film's journey from the sensor to the camera position using the formula above, the tension fluctuations generated in the past can be accurately mapped onto the currently observed thin film image patch. This is a prerequisite for root cause analysis.
[0025] This step involves carefully selecting three core mechanical parameters: tension value, main line velocity, and displacement of the correction mechanism. It also constructs a spatiotemporal feature correlation mechanism based on linear velocity adaptation. This achieves the effect of covering the core causes of wrinkle defects with minimal data and significantly reducing the complexity of multimodal models. By introducing a compensation formula, it achieves the effect of accurately aligning visual images with equipment status time-series data, eliminating erroneous attributions caused by system errors, and ensuring the accuracy of correction commands.
[0026] Step S200: Run a lightweight perception algorithm based on reparameterized convolution to perform a sliding window scan on the surface image. When the gradient change rate of the detected region exceeds the dynamic threshold, it is determined to be a suspected defect, triggering a high-frequency sampling mechanism to obtain the image of the center region of the suspected defect and the sequence of process parameter changes within a set time window before and after the defect occurs. This step aims to solve the problem that traditional deep models cannot meet the real-time detection requirements of high-speed roll material movement under high-resolution images.
[0027] For example, a lightweight perceptron algorithm model network based on EffiRepCSANet is constructed. For instance, during the training phase, a network containing... convolution A multi-branch structure involving convolution and residual branches; during the inference phase, operator fusion is performed using formulas, manifested as follows: In the formula, the and Weights for convolution kernels of different sizes; It is the identity matrix, corresponding to the residual branch.
[0028] After reparameterization, the single-path convolutional layer performs sliding window feature extraction on the image stream. When the gradient change rate of pixels in a region exceeds a dynamic threshold, it is identified as a suspected defect. This step utilizes reparameterization technology to compress the complex multi-path training network into an efficient single-path matrix inference network, which can significantly improve inference speed and reduce memory usage without sacrificing accuracy.
[0029] Step S300: Input the image of the suspected defect center region and the sequence of process parameter changes within a set time window T seconds before and after the defect occurrence into the multimodal deep learning model to obtain image feature branch vectors respectively. and process timing feature branch vector The model then uses a cross-attention attribution fusion model to output the wrinkle type C, mask M, and root cause attribution score A. This step aims to address the limitations of a single modality, namely, the difficulty in distinguishing the physical causes of wrinkles from images alone, and the challenge that fine wrinkles and large bulges are extremely unevenly distributed in the feature space, making it difficult for ordinary convolution to capture them simultaneously.
[0030] Further, step S300 includes (see...) Figure 3 ): Step S301: Input the image of the suspected defect center region into the adaptive high-low frequency fusion AHLF module. The high-frequency branch, i.e., sub-branch 1 (Hi-Fi Branch), uses a local window attention model to capture wrinkle edge details, while the low-frequency branch, i.e., sub-branch 2 (Lo-Fi Branch), models the overall deformation of the thin film, i.e., large-area bulges, through global pooling, and finally obtains the image feature branch vector. By leveraging the characteristics of the AHLF module, a pure visual feature vector is extracted that includes both wrinkle edge details (high frequency) and overall film deformation (low frequency). This forces the identified wrinkles to conform to physical laws, shielding pure optical interference at the mathematical model level. This avoids the problem of missed detection when wrinkles of different sizes coexist in the same image, and solves the difficulty of wrinkle recognition caused by the transparency, reflectivity, and multi-scale nature of packaging films in dynamic roll packaging scenarios.
[0031] In a preferred embodiment of the present invention, specifically for the application scenario of a high-reflectivity aluminum foil packaging line, when the camera captures localized strong reflections caused by ambient light, the reflective component manifests as high-frequency, high-brightness pulse noise in the feature space. At this time, the AHLF module performs feature filtering: the high-frequency branch calculates the second-order gradient direction feature map within the local window. Since the reflective point lacks the linear extension characteristic unique to wrinkles, its response value in the high-frequency feature map is initially suppressed; the low-frequency branch obtains the deformation envelope of the current film through global pooling operations. The final fused feature map retains only the region that simultaneously possesses edge texture and surface deformation features. This embodiment reduces the false alarm rate of micro-wrinkles from 15.2% in the traditional algorithm to below 0.8%, achieving stable identification of sub-millimeter-level wrinkles.
[0032] Step S302: Set the tension value within a time window before and after the defect occurs. Main line speed and displacement of the correction mechanism The sequence is input into a one-dimensional temporal convolutional network to extract the process time sequence feature branch vector. .
[0033] Step S303: Convert the image feature branch vector and process timing feature branch vector The input is fed into the cross-attention attribution fusion model to obtain the wrinkle type C, segmentation mask M, and root cause attribution score A.
[0034] Furthermore, the cross-attention attribution fusion model Image feature branch vector For the query term, the process time sequence feature branch vector The key-value items are fused to calculate the contribution of process fluctuations to wrinkles at specific locations. The relationship is expressed as follows: in For feature dimensions.
[0035] Further, step S303 includes: Step S3031: Through the cross-attention attribution fusion model A multimodal fusion feature matrix is obtained. The multimodal fusion feature matrix After global average pooling, the input is a linear classification layer. By calculating the probability distribution of each category, the defect category C, i.e., the classification head, is output; for example, longitudinal wrinkles, transverse wrinkles, and local hollow areas.
[0036] Step S3032: Use a deconvolutional network to process the multimodal fusion feature matrix. Spatial dimension restoration is performed to generate a binary mask M of the same size as the original image, thereby realizing defect localization, i.e., the segmentation head; each pixel in the image is scored, and the high-scoring point is the wrinkled region, forming a segmentation mask, and outputting a pixel-level defect segmentation map.
[0037] Step S3033: Extract the cross-attention attribution fusion model The weighted coefficient matrix is used to output the root cause attribution score A by statistically analyzing the contribution rate of each process parameter feature channel to the salient region of the image. By matching the wrinkles in the image with the corresponding process features, if the tension channel has the highest weight, the attribution score A points to tension abnormality. This solves the black box problem of deep learning and can directly locate the physical cause of the defect.
[0038] This step achieves more accurate results than pure visual detection by deeply fusing real-time process parameters and visual features in a multimodal network. By analyzing the weights of the cross-attention mechanism within the model, the contribution of each process parameter to the defect is quantified and an attribution score is output, thereby giving the deep learning model interpretability and providing clear and operable correction guidance for the control system.
[0039] Step S400: Based on the wrinkle type C, mask M, and root cause score A, calculate the quantitative correction increment value for the tension regulating valve or the correction actuator under the current defect state.
[0040] Further, step S400 includes: Step S401: Based on the wrinkle type C and the mask M, calculate the severity index S of the current defect. The calculation formula is as follows: In the formula, It is the total number of defective pixels in mask M. It is the total number of pixels in the image. These are the weighting coefficients for different categories of defects. For example, longitudinal wrinkles are more harmful, so they have a higher weight. It is the average gradient deviation of the defect area.
[0041] Step S402: Calculate the control increment under multimodal guidance by combining the root cause attribution score A. The calculation formula is: It locks the sensitivity gain of each of the actuators. This is a direction determination function, where i represents the actuator, i=1 is the tension regulating valve, and i=2 is the correction actuator; if attributed to tension... If it is too large, a tension reduction command will be generated; if attributed to phase... If there is an offset, a correction position adjustment amount will be generated.
[0042] This step, through the two formulas mentioned above, achieves an objective and quantitative assessment of the severity of defects, enabling precise matching of adjustment amounts to the root causes of defects, avoiding blind or equal adjustments to each actuator, shortening defect correction time, and improving control efficiency.
[0043] Step S500: Send the correction position adjustment amount to the corresponding actuator for execution.
[0044] See Figure 2 The present invention also provides a technical solution: a method for identifying and correcting wrinkle defects in roll packaging based on visual feedback, characterized in that it includes the following system: The system includes a perception module, a high-speed recognition module, a precise diagnosis module, a decision control module, and a correction execution module, wherein... The sensing module, including a camera, a light source, and a sensor gateway, is configured to synchronously acquire images of the thin film surface and real-time process parameters, and perform spatiotemporal alignment processing on the data. For example, the light source of the sensing module adopts a low-angle grazing strip light source array to enhance the contrast of the wrinkle features on the thin film surface.
[0045] The high-speed recognition module is communicatively connected to the perception module and is configured to run a lightweight, heavily parameterized network to monitor the image stream in real time and trigger a warning of suspected defects. The precise diagnosis module is communicatively connected to the high-speed identification module and is configured to deploy a multimodal deep learning model for classifying, segmenting, and performing root cause analysis based on cross-attention for suspected defects. The decision control module is communicatively connected to the precise diagnosis module and is configured to calculate and generate quantitative adjustment instructions for each actuator based on the diagnosis results. The correction execution module, together with the decision control module, including the tension adjustment mechanism and / or the correction execution mechanism, is configured to receive and execute the quantitative adjustment command.
[0046] An electronic device includes a processor and a memory, characterized in that the processor runs a computer program or code stored in the memory to implement the visual feedback-based method for identifying and correcting wrinkles in roll packaging.
[0047] A computer-readable storage medium for storing computer programs or code, characterized in that, when the computer program or code is executed by a processor, it implements the method for identifying and correcting wrinkles in roll-type packaging based on visual feedback.
[0048] This invention constructs a fully closed-loop intelligent control system integrating perception, diagnosis, decision-making, execution, and verification. This system achieves automated and precise control of wrinkle defects during the packaging process. Utilizing a linear velocity-adaptive spatiotemporal feature association mechanism, it accurately aligns visual images with real-time process parameters such as tension and displacement, eliminating attribution bias caused by physical spatial lag at the source. Subsequently, a lightweight perception network based on heavily parameterized convolution achieves millisecond-level initial screening of suspected defects. A multimodal deep learning model integrating AHLF adaptive high- and low-frequency feature extraction and cross-attention mechanisms is introduced. This not only achieves pixel-level wrinkle segmentation and localization but also outputs attribution scores by quantifying the weights of each process parameter's influence on features, giving the model interpretability in analyzing the physical causes of defects. Finally, the system combines quantitative correction commands to drive the actuator's actions. This solution improves the system's real-time response capability and recognition reliability in high-speed, complex environments, overcoming the shortcomings of blind and lagging manual experience adjustments, unclear attribution in traditional pure visual detection, and slow response in deep learning visual detection models. This significantly reduces packaging material waste and improves the intelligent production efficiency of the production line.
[0049] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.
[0050] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0051] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A method for identifying and correcting wrinkle defects in rolled packaging based on visual feedback, characterized in that, Includes the following steps: Step S100: During the roll wrapping process, simultaneously acquire real-time images of the film surface. And process feature parameters, using a unified timestamp to spatiotemporally align the surface image with the process feature parameters; Step S200: Run a lightweight perception algorithm based on reparameterized convolution to perform sliding window scanning on the surface image. When the gradient change rate of the detected region exceeds the dynamic threshold, it is determined to be a suspected defect, triggering a high-frequency sampling mechanism to obtain the image of the center region of the suspected defect and the sequence of process parameter changes within a set time window before and after the defect occurs. Step S300: Input the image of the suspected defect center region and the sequence of process parameter changes within a set time window T seconds before and after the defect occurrence into the multimodal deep learning model to obtain image feature branch vectors respectively. and process timing feature branch vector And through the cross-attention attribution fusion model, the final output is wrinkle type C, mask M and root cause attribution score A; Step S400: Based on the wrinkle type C, mask M, and root cause score A, calculate the quantitative correction increment value for the tension regulating valve or the correction actuator under the current defect state; S500: The correction position adjustment amount is sent to the corresponding actuator for execution.
2. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 1, characterized in that, Step S100 further includes: Step S101: During the roll wrapping process, an industrial camera captures real-time images of the film surface. The programmable logic controller (PLC) acquires process characteristic parameters in real time, and the same pulse signal simultaneously triggers camera exposure and PLC acquisition. The process characteristic parameters include the tension value detected by the tension sensor. Main line speed and displacement of the correction mechanism ; Step S102: Based on image frames The precise exposure time is Find the data in the PLC that matches Two adjacent time points and After time alignment, the tension value corresponding to that frame is... The calculation formula is: ； Step S103: Perform spatial correction based on the physical distance D between the industrial camera and the tension sensor, then the tension value after frame alignment is obtained. In the formula To compensate for the time difference in the travel time of the thin film from the tension sensor to the camera position, .
3. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 1, characterized in that, Step S300 further includes: Step S301: Input the image of the suspected defect center region into the adaptive high-low frequency fusion AHLF module, wherein the high-frequency branch uses a local window attention model to capture wrinkle edge details, and the low-frequency branch uses global pooling to model the overall deformation of the thin film to obtain the image feature branch vector. ; Step S302: Set the tension value within a time window before and after the defect occurs. Main line speed and the displacement of the correction mechanism The sequence is input into a one-dimensional temporal convolutional network to extract the process time sequence feature branch vector. ; Step S303: Convert the image feature branch vector and process timing feature branch vector The input is fed into the cross-attention attribution fusion model to obtain the wrinkle type C, segmentation mask M, and root cause attribution score A.
4. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 3, characterized in that: The cross-attention attribution fusion model Image feature branch vector For the query term, the process time sequence feature branch vector The key-value items are fused to calculate the contribution of process fluctuations to wrinkles at specific locations. The relationship is expressed as follows: in For feature dimensions.
5. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 4, characterized in that, Step S303 further includes: Step S3031: Through the cross-attention attribution fusion model A multimodal fusion feature matrix is obtained. The multimodal fusion feature matrix After performing global average pooling, the input is a linear classification layer. By calculating the probability distribution of each category, the defect category C is output. Step S3032: Use a deconvolutional network to process the multimodal fusion feature matrix. Perform spatial dimension restoration, generate a binarized mask M of the same size as the original image, and output a pixel-level defect segmentation map; Step S3033: Extract the cross-attention attribution fusion model The weight coefficient matrix is calculated, and the root cause attribution score A is output by statistically analyzing the contribution rate of each process parameter feature channel to the salient region of the image.
6. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 1, characterized in that, Step S400 further includes: Step S401: Calculate the severity index S of the current defect based on the wrinkle type C and the mask M; Step S402: Calculate the control increment under multimodal guidance by combining the root cause attribution score A. The calculation formula is: Where i represents the actuator, i=1 is the tension regulating valve, i=2 is the correction actuator; It locks the sensitivity gain of each of the actuators. It is a direction determination function.
7. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 6, characterized in that, Including the following systems: The system includes a perception module, a high-speed recognition module, a precise diagnosis module, a decision control module, and a correction execution module, wherein... The sensing module, including a camera, a light source, and a sensor gateway, is configured to simultaneously acquire thin film surface images and real-time process parameters, and perform spatiotemporal alignment processing on the data. The high-speed recognition module is communicatively connected to the perception module and is configured to run a lightweight, heavily parameterized network to monitor the image stream in real time and trigger a warning of suspected defects. The precise diagnosis module is communicatively connected to the high-speed identification module and is configured to deploy a multimodal deep learning model for classifying, segmenting, and performing root cause analysis based on cross-attention for suspected defects. The decision control module is communicatively connected to the precise diagnosis module and is configured to calculate and generate quantitative adjustment instructions for each actuator based on the diagnosis results. The correction execution module, together with the decision control module, including the tension adjustment mechanism and / or the correction execution mechanism, is configured to receive and execute the quantitative adjustment command.
8. The method for identifying and correcting wrinkle defects in roll packaging based on visual feedback according to claim 7, characterized in that: The sensing module uses a low-angle grazing strip light source array to enhance the contrast of the wrinkle features on the thin film surface.
9. An electronic device comprising a processor and a memory, characterized in that, The processor runs a computer program or code stored in the memory to implement the visual feedback-based method for identifying and correcting wrinkles in roll packaging as described in any one of claims 1 to 8.
10. A computer-readable storage medium for storing computer programs or code, characterized in that, When the computer program or code is executed by a processor, it implements the method for identifying and correcting wrinkle defects in roll packaging based on visual feedback as described in any one of claims 1 to 8.