A production process closed loop optimization method based on PVC floor film surface defect detection

Abstract

The application discloses a production process closed-loop optimization method based on PVC floor film surface defect detection, and relates to the technical field of intelligent manufacturing surface detection. Through adaptive adjustment of light source angle and camera exposure parameters of a dynamic structure light source and a linear array camera, a clear and tiny defect image is acquired; based on a BRDF physical imaging model, specular reflection and diffuse reflection components are separated, a defect recognition model of feature space reconstruction and dynamic boundary constraint is adopted, pixel-level reconstruction error is calculated to generate a defect atlas; defect distribution is extracted by using space-time correlation analysis to generate defect thermal data; a mapping relationship between defect characteristics and process variables is established through a process parameter deviation tracing model to inversely deduce deviation types and degrees; process parameters are adjusted by issuing instructions under phased PID control to realize adaptive optimization. The application effectively suppresses imaging interference, improves the detection rate of tiny defects, quickly locates defect sources and automatically corrects process parameters, and is suitable for high-speed and high-reflectivity PVC floor film continuous production lines.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing surface inspection technology, specifically to a closed-loop optimization method for production processes based on the detection of surface defects in PVC flooring films. Background Technology

[0002] As a core material for home and commercial floor decoration, the surface appearance quality of PVC flooring directly determines product grade, market pricing, and user experience. On high-speed continuous production lines integrating extrusion, calendering, lamination, and printing, micron-level defects such as crystal points, bubbles, scratches, color differences, pitting, foreign objects, and wrinkles are easily generated. Traditional manual inspection suffers from low speed, fatigue, strong subjectivity, high missed detection rate, and lack of full-process traceability, making it difficult to meet the zero-defect and high-consistency quality requirements of high-end building materials. With the deepening of intelligent manufacturing and Industry 4.0, online defect detection based on machine vision is gradually replacing manual labor and becoming the mainstream direction for quality control in the industry. However, how to achieve high-precision detection of minute defects under high-speed, high-reflectivity, and complex texture conditions, and how to convert the detection results into process adjustment instructions in real time to form a closed-loop optimization, remains a key bottleneck restricting the improvement of quality and efficiency of PVC flooring production lines.

[0003] Currently, most mainstream visual inspection systems in the industry use fixed light sources and fixed exposure parameters. When production line speed fluctuates, film texture changes, or ambient light changes, problems such as motion blur, insufficient defect contrast, and specular reflection interference easily occur, leading to missed detection of minor defects or false alarms. Most systems only complete defect identification, marking, and alarming, without establishing a correlation mapping between defect features and production process parameters. They cannot automatically determine whether the defect originates from extrusion temperature, screw speed, traction tension, calender roll gap, or cooling air temperature fluctuations. Inspection data and production control are disconnected, relying solely on manual experience for post-production adjustments, easily resulting in batches of scrap and raw material waste. Existing equipment generally lacks integrated mechanisms such as adaptive imaging, dynamic boundary constraint anomaly detection, spatiotemporal correlation tracing, and staged PID closed-loop regulation. The detection accuracy, response speed, and intelligence level are insufficient to meet the demands of large-scale, stable production of high-end PVC flooring films.

[0004] Extensive research has been conducted in academia on visual inspection of surface defects in thin-film materials, providing theoretical support for industrial applications. In their 2020 paper, *Automatic surface inspection for SPVC using a composite vision-based method*, Qilin Bi et al. proposed a composite vision inspection framework for rigid PVC products. By optimizing the imaging system and region-of-interest extraction method, they reduced background interference and improved real-time detection performance. Experiments demonstrated the advantages of this approach in terms of low latency and robustness. A 2023 article in *Industrial Engineering*, titled "Research Progress in Surface Defect Detection Based on Deep Intelligent Vision," systematically reviewed the applications of deep learning in defect classification, localization, and segmentation, pointing out that multi-task fusion and lightweight models are key paths to improving applicability in industrial scenarios. A 2025 review in *Journal of Image and Graphics*, titled "Vision-Based Incompletely Annotated Surface Defect Detection," addressed the high cost and imbalanced sample problem of industrial defect annotation, summarizing weakly supervised and semi-supervised learning methods to provide solutions for training small-sample defect models. The aforementioned literature advances the development of detection technology from the perspectives of optical modeling, algorithm architecture, and sample learning, but all focus on offline or semi-online detection processes. They do not cover the entire chain of methods, including adaptive imaging, dynamic constraint of feature space, spatiotemporal tracing of defects, and real-time closed-loop control of production processes, making it difficult to directly support the high-precision intelligent feedback requirements of high-speed PVC flooring film production lines.

[0005] Chinese patent CN110068578B (A method, apparatus, and terminal device for detecting surface defects in PVC flooring) uses multi-view, variable-light imaging to acquire images of the PVC flooring surface. After preprocessing, segmentation, feature extraction, and using a preset model, defect identification and classification are completed, improving the ability to detect complex defects to a certain extent. Chinese patent CN116101708A (A device and method for detecting surface defects in PVC flooring) designs a dedicated detection device for flooring surface defects, optimizing the image acquisition and defect discrimination process and improving detection stability. Chinese patent CN120510088A (A method and related apparatus for detecting surface defects in thin films) introduces the fusion of point cloud data and dynamic image features to improve the accuracy of defect detail extraction. While improvements have been made in imaging methods, defect identification, and device structure, significant shortcomings remain: the light source angle and camera exposure have not been adaptively adjusted according to line speed and texture; an anomaly detection mechanism based on feature space reconstruction and dynamic boundary constraints has not been adopted, resulting in insufficient robustness to small, low-contrast defects; periodic and clustered root cause tracing has not been performed based on the spatiotemporal distribution of defects; defect thermal data has not been spatially correlated with process parameters, and a phased PID closed-loop optimization has not been formed. Detection and production remain in an open-loop mode of "detection-alarm-manual adjustment," failing to suppress defect generation at its source. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a closed-loop optimization method for a production process based on the detection of surface defects in PVC flooring film, specifically including: Image data of PVC floor membrane surface is collected. Based on the real-time running speed and surface texture type of PVC floor membrane, the incident angle of the light source and the exposure parameters of the camera are adaptively adjusted by a dynamic structured light source and a line scan camera to obtain image data of minute surface defects under different working conditions. The collected surface image data is preprocessed. An ambient light interference suppression model is established by using an adaptive image enhancement algorithm based on the physical imaging model and combining it with the optical reflection characteristics of PVC material. By separating the specular reflection component and the diffuse reflection component, the texture and morphology features of the defect area are obtained, and a set of feature vectors to be tested is constructed. The set of feature vectors to be tested is input into a pre-trained defect recognition model. The defect recognition model adopts an anomaly detection architecture based on feature space reconstruction. By introducing a dynamic boundary constraint mechanism in the normal sample feature space, the pixel-level reconstruction error between the feature vectors to be tested and the feature distribution of normal samples is calculated, and a defect map is generated. Based on the defect map, the distribution of defects in the width and length directions of the PVC floor film is extracted using a spatiotemporal correlation analysis algorithm. By analyzing the periodicity, clustering, and directionality of the defect distribution, the potential root causes of the defects are identified, and thermal data of the defect distribution is generated. The defect distribution thermal data is correlated and mapped with the production process parameter space to construct a process parameter deviation tracing model. By analyzing the correlation between defect distribution characteristics and historical data of process variables, the type and degree of process parameter deviation that caused the defect are inferred, and control instructions are generated. The control commands are sent to the production line control system in real time, and the production process parameters of the PVC floor film are adjusted in stages to achieve closed-loop adaptive optimization of the production process.

[0007] Preferably, the step of adaptively adjusting the incident angle of the light source and the camera exposure parameters by means of a dynamic structured light source and a line scan camera according to the real-time running speed and surface texture type of the PVC floor membrane includes: Real-time acquisition of production line operating speed and preset PVC floor membrane surface texture type ; According to the running speed Calculate the camera's exposure time This makes the motion blur radius Less than the preset fuzzy threshold ,in , The magnification of the optical system; According to the surface texture type Query the preset illumination angle mapping table to determine the target incident angle of the structured light source. The illumination angle mapping table stores the correspondence between different texture types and the optimal incident angle, which is used to enhance the defect contrast under a specific texture.

[0008] Preferably, the step of adjusting the running speed Calculate the camera's exposure time It also includes: Obtain the camera's signal-to-noise ratio With exposure time nonlinear response function ,in Light source intensity; In satisfying Determine the optimal exposure time under the constraints. : ,in, and These are weighting coefficients, used to balance the trade-off between image signal-to-noise ratio and motion blur.

[0009] Preferably, the adaptive image enhancement algorithm based on the physical imaging model, combined with the optical reflectivity of PVC material, establishes an ambient light interference suppression model, including: Construct a physical imaging model based on the bidirectional reflectance distribution function (BRDF) to measure the intensity of the acquired images. Represented as ambient light component Specular reflection component and diffuse reflection component The linear superposition is given by the formula: ; Multiple frames of images are acquired using a polarization camera or multi-angle light source, and the specular reflection component is separated using the polarization difference method or photometric stereo method. and diffuse reflection component ; Ambient light component Background modeling and subtraction were performed to obtain a defect feature image free from ambient light interference. The formula is: ,in This represents the average grayscale value of the background area.

[0010] Preferably, the process involves acquiring multiple frames of images using a polarization camera or multi-angle light source, and then separating the specular reflection component using the polarization difference method. and diffuse reflection component ,include: Controlling the polarization camera in two orthogonal polarization directions and Images were acquired separately below and ; Calculate the degree of polarization based on Fresnel reflection principle. and polarization angle polarization degree Used to characterize the intensity of specular reflection: Based on the degree of polarization Construct a specular reflection mask Regions with polarization degrees higher than a preset threshold are identified as specular reflection regions, thereby separating them. .

[0011] Preferably, the defect identification model adopts an anomaly detection architecture based on feature space reconstruction, which introduces a dynamic boundary constraint mechanism in the feature space of normal samples, including: Construct a feature extraction network based on an autoencoder and a decoding network Input image Mapping to feature space Then reconstructed into an image ; In feature space In the middle, define a dynamic boundary sphere. ,in The center of the normal sample features, The boundary radius; During the training phase, minimize the reconstruction error of normal samples. At the same time, constrain the characteristics of normal samples Falling inside the boundary sphere means minimizing the boundary constraint loss. ; During the inference phase, the features of the test sample are calculated. To the center distance ,like If so, it is judged as abnormal.

[0012] Preferably, the step of calculating the pixel-level reconstruction error between the feature to be tested and the feature distribution of normal samples to generate a defect map includes: Computational reconstruction of the image Compared with the original input image Pixel-level difference map ; Calculate the features to be measured Anomaly score in feature space ,in The standard deviation of the characteristics of normal samples; pixel-level difference map With feature-level anomaly scores Weighted fusion is performed to generate a defect confidence map. The formula is: ,in, To integrate weights, This is an operation that upsamples the feature map size to the original image size.

[0013] Preferably, the step of extracting the distribution of defects in the width and length directions of the PVC flooring membrane using a spatiotemporal correlation analysis algorithm includes: By arranging continuous defect maps in a time series, a spatiotemporal distribution matrix of defects is constructed. The rows of the matrix correspond to the width coordinates of the production line, and the columns correspond to the time steps. For the spatiotemporal distribution matrix of the defects Perform a two-dimensional Fourier transform to obtain the spectrum. ; Analyze the spectrum peak position ,in The corresponding periodic frequency of the defect along its length. The periodic frequency of the corresponding defect in the width direction; According to the periodic frequency Calculate the period length of the defect , To determine the operating speed of the production line, identify periodic defects associated with the traction roller or embossing roller.

[0014] Preferably, the construction of the process parameter deviation tracing model, by analyzing the correlation between defect distribution characteristics and historical data of process variables, reverse-engineers the type and degree of process parameter deviations that lead to defects, including: Construct a multi-input, multi-output regression model, with the input being the defect distribution feature vector. The output is a vector of process parameter deviations. ; The regression model is trained using historical production data to learn the mapping relationship from defect characteristics to process deviations; During the online detection phase, the defect distribution feature vector is extracted in real time. Input the trained model to predict the deviation of current process parameters. .

[0015] Preferably, the step of sending control commands to the production line control system in real time and adjusting the production process parameters of the PVC flooring film in stages includes: Predicted process parameter deviations This is converted into a control command, which includes adjusting the target value. ; A phased PID control strategy is adopted, dividing the adjustment process into two stages: coarse adjustment and fine adjustment. In the coarse adjustment phase, a larger scaling factor is used. Rapidly approach the target value; when the error between the process parameters and the target value is less than the preset threshold, switch to the fine-tuning stage and use a smaller scaling factor. and integral coefficient To eliminate steady-state error, the control law is: ,in, This represents the error in the process parameters at the current moment. This is the current adjustment phase.

[0016] Compared with the prior art, the technical solution of this application has the following technical effects: This invention, through the adaptive and coordinated adjustment of a dynamic structured light source and a line scan camera, can stably acquire clear images of the PVC floor membrane surface under different operating speeds and texture types, effectively suppressing interference caused by ambient light and material reflection, and fully presenting the texture and morphology information of minute defects, providing a reliable data foundation for subsequent defect identification.

[0017] This invention employs a defect identification model based on feature space reconstruction and dynamic boundary constraints, which can accurately extract defect features and calculate pixel-level reconstruction errors, stably generate defect maps, accurately distinguish between real defects and interference information, improve the integrity and reliability of defect detection, and ensure the consistency of defect location and type determination.

[0018] This invention utilizes a spatiotemporal correlation analysis algorithm to extract defect distribution features, effectively identifying the root causes and formation patterns of defects, generating intuitive thermal data on defect distribution, and enabling rapid location and quantitative presentation of defect causes, providing a clear and practical basis for adjusting process parameters.

[0019] This invention maps defect data with production process parameters to form closed-loop control commands. By adaptively adjusting process parameters in stages, it continuously optimizes the production process, steadily improves product surface quality, and achieves fully automated operation of the entire process of detection, analysis, traceability, and control, ensuring the long-term stable and efficient operation of the production system.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0023] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows: Figure 1 Flowchart of the overall process optimization method for production process closed-loop optimization based on PVC flooring film surface defect detection; Figure 2 A schematic diagram illustrating the principle of image preprocessing and optical component separation on the surface of PVC flooring film; Figure 3 Architecture diagram of a defect identification model based on feature space reconstruction and dynamic boundary constraints; Figure 4 A schematic diagram of the process for generating spatiotemporal correlation analysis and distributed thermal data of defects; Figure 5 A comparison curve of signal-to-noise ratio changes of different defect detection schemes under minute defects; Figure 6 Scatter plot comparing the spatial distribution of defect features of the present invention and the prior art; Figure 7 Scatter plot comparing the accuracy of defect cycle calculation for each scheme under harsh working conditions; Figure 8 A comparative diagram of the defect identification effects of different methods under strong reflective interference; Figure 9 A comparison curve of the convergence trajectory and control stability during the closed-loop adjustment of process parameters. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0025] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0026] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0027] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0028] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0029] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion. Example 1

[0030] This embodiment mainly describes a closed-loop optimization method for the production process based on the detection of surface defects in PVC flooring film, such as... Figure 1 As shown, it specifically includes: Image data of PVC floor membrane surface is collected. Based on the real-time running speed and surface texture type of PVC floor membrane, the incident angle of the light source and the exposure parameters of the camera are adaptively adjusted by a dynamic structured light source and a line scan camera to obtain image data of minute surface defects under different working conditions. The collected surface image data is preprocessed. An ambient light interference suppression model is established by using an adaptive image enhancement algorithm based on the physical imaging model and combining it with the optical reflection characteristics of PVC material. By separating the specular reflection component and the diffuse reflection component, the texture and morphology features of the defect area are obtained, and a set of feature vectors to be tested is constructed. The set of feature vectors to be tested is input into a pre-trained defect recognition model. The defect recognition model adopts an anomaly detection architecture based on feature space reconstruction. By introducing a dynamic boundary constraint mechanism in the normal sample feature space, the pixel-level reconstruction error between the feature vectors to be tested and the feature distribution of normal samples is calculated, and a defect map is generated. Based on the defect map, the distribution of defects in the width and length directions of the PVC floor film is extracted using a spatiotemporal correlation analysis algorithm. By analyzing the periodicity, clustering, and directionality of the defect distribution, the potential root causes of the defects are identified, and thermal data of the defect distribution is generated. The defect distribution thermodynamic data is correlated and mapped with the production process parameter space to construct a process parameter deviation tracing model. By analyzing the correlation between defect distribution characteristics and historical data of process variables, the type and degree of process parameters that cause defects are deduced, and control commands are generated. The control commands are sent to the production line control system in real time, and the production process parameters of the PVC floor film are adjusted in stages to achieve closed-loop adaptive optimization of the production process.

[0031] Furthermore, the acquisition module acquires surface image data with no motion blur, high signal-to-noise ratio, and optimal defect contrast during continuous operation of the production line. The acquisition module includes a running speed acquisition unit, a texture type recognition unit, an exposure parameter calculation unit, and a light source angle adjustment unit; it also acquires the running speed signal of the production line traction mechanism in real time. The speed acquisition range is 0 m / min to 120 m / min, the acquisition accuracy is ±0.1 m / min, and the data update frequency is 100 Hz; texture type Based on the production formula, it is divided into four types: smooth flat surface, fine textured surface, deep embossed surface, and synchronous matching textured surface. Each type corresponds to a fixed combination of optical reflection characteristics and optimal imaging parameters. The motion blur radius is determined by the linear velocity, exposure time, and optical magnification, as shown in the following formula: In the formula, The motion blur radius is expressed in pixels. The camera exposure time ranges from 10 μs to 1000 μs. This is the magnification of the optical system, with a fixed value of 0.025, representing 25 image pixels per millimeter of physical length.

[0032] Preset fuzzy threshold The value is fixed at 1 pixel, and all exposure parameter calculations must meet the following constraints. This ensures that the edge contours of defects are not blurred, blurred, or deformed, thus guaranteeing the morphological integrity of minor defects. Under the premise of satisfying motion blur constraints, the optimal exposure time is solved by maximizing the image signal-to-noise ratio (SNR). (Camera SNR) With exposure time Light source intensity There exists a fixed nonlinear response relationship, the formula of which is: Among them, the light source intensity Adjustable from 0 to 255 levels, corresponding to 0-100% power output. Optimal exposure time. The method of finding the extreme value with constraints is used, and the specific formula is as follows: In the formula, This is the signal-to-noise ratio weighting coefficient, with a fixed value of 0.7; The fuzzy penalty weight coefficient is fixed at 0.3; the weight coefficient satisfies the normalization constraint. This allows the system to maintain an optimal balance between signal-to-noise ratio and motion blur suppression.

[0033] The incident angle of the light source depends on the texture type. Adaptive adjustment; the system has a built-in illumination angle mapping table, and the smooth plane corresponds to the incident angle. Subtle textures face Deep embossing corresponds to Synchronously match the pattern surface The angle adjustment accuracy is ±0.5°, and the adjustment range covers 15° to 75°. By maximizing the grayscale difference between defects and background texture through optimal incident angle, interference from the texture itself is suppressed, and the detectability of minute defects such as crystal points, bubbles, scratches, and pits is improved.

[0034] Furthermore, such as Figure 2 As shown, a physical imaging model is constructed based on the bidirectional reflectance distribution function (BRDF) to achieve complete separation of ambient light components, specular reflection components, and diffuse reflection components, eliminating interference and preserving the true features of defects; the image is displayed in two-dimensional plane coordinates. Strength at the point It can be linearly decomposed into three independent components, as shown in the following formula: In the formula, The original image grayscale value, ranging from 0 to 255; For ambient light component, there is a fixed grayscale offset across the entire area, with a value range of 10 to 40. The specular reflection component is generated by the smooth surface of PVC material and appears as a localized bright area with a grayscale value usually higher than 200. The diffuse component contains the material's texture and true information about defects, and is the basis for defect feature extraction.

[0035] Component separation is achieved using the polarization difference method. The system controls a polarization camera to simultaneously acquire four frames of images in four orthogonal directions: 0°, 45°, 90°, and 135°, denoted as _____. , , , Calculation of total polarization degree based on Fresnel reflection principle The calculation formula is as follows: In the formula, The polarization degree is normalized to a value between 0 and 1; a higher value indicates a higher specular reflection intensity. The system sets a fixed polarization degree threshold of 0.7. The region is identified as a specular reflection region, and a binary specular reflection mask is constructed. To achieve the specular reflection component Precise separation and elimination.

[0036] Ambient light component The global Gaussian background modeling method is used for estimation. The average gray value of the background region is calculated by selecting 100 consecutive rows of defect-free image data. , The calculation window size is 64×64 pixels, and the calculation results are updated in real time. After subtracting the ambient light mean from the diffuse reflection component, an interference-free defect feature image is obtained. The calculation formula is as follows: The image is free from ambient light interference, specular highlights, and texture overlay, preserving the complete texture and shape features of defects. Based on this image, multi-dimensional features such as gray-level mean, gradient magnitude, directional gradient, local contrast, gray-level variance, texture energy, texture entropy, and shape moments are extracted to construct a fixed-dimensional feature vector set. The vector dimension is uniformly 256, and each feature value is normalized to the range of 0 to 1 to meet the input requirements of the defect recognition model.

[0037] Furthermore, such as Figure 3As shown, the defect identification model adopts an anomaly detection architecture with an autoencoder structure, consisting of a feature extraction network. With reconstructed decoding network The network consists of fully convolutional layers, including convolutional layers, batch normalization layers, activation layers, max pooling layers, and upsampling layers. The input image size is uniformly 512×512 pixels with 1 channel. The feature extraction network has 8 convolutional layers with kernel sizes of 3×3, 3×3, 5×5, 5×5, 7×7, 7×7, 9×9, and 9×9, outputting a low-dimensional, compact feature vector. The dimension is 64. The reconstruction decoding network is symmetrical with the feature extraction network, with a total of 8 deconvolution and upsampling layers, which restores the low-dimensional features to a 512×512 pixel reconstructed image.

[0038] The mapping relationship from image to feature space is as follows: The mapping relationship from the feature space to the reconstructed image is as follows: In the formula, To input the defect feature image, Output the reconstructed image to the model. These are the hidden layer feature vectors.

[0039] Constructing a dynamic boundary sphere within the feature space This is used to define the distribution range of normal samples. Dynamic boundary sphere center. The radius of the dynamic boundary sphere is calculated by averaging the feature vectors of all normal samples, with the calculation window being the most recent 10,000 frames of normal samples and an update frequency of 10Hz. From normal sample features to center The maximum Euclidean distance is determined by the following formula: ,in For the first A normal sample feature, radius It adaptively adjusts in real time as normal samples are updated, ensuring that the boundary covers the entire distribution area of ​​normal samples.

[0040] During model training, reconstruction error and boundary constraint error are optimized simultaneously. The total loss function consists of a weighted average of these two parts. The reconstruction loss measures the pixel-level difference between the input image and the reconstructed image, and is calculated using the mean squared error, as shown in the following formula: Boundary constraint loss is used to ensure that all normal sample features fall inside the dynamic boundary sphere, and the formula is as follows: The total loss function is: In the formula, This is the boundary constraint coefficient, with a fixed value of 0.2, used to balance the reconstruction error and the boundary constraint strength.

[0041] During the inference phase, the features of the sample to be tested are denoted as... Calculate its distance to the center Euclidean distance: like If it is determined to be an abnormal defect area; It was determined to be a normal area.

[0042] Simultaneously, both pixel-level and feature-level anomaly scores are calculated. The pixel-level difference map directly reflects the degree of local anomalies, as shown in the following formula: Feature-level anomaly scores reflect the degree of anomaly in the overall distribution. They are calculated based on the Gaussian kernel function, as shown in the following formula: In the formula, The global standard deviation of the normal sample characteristics is obtained from the statistics of historical normal samples and is fixed at 0.15.

[0043] Feature-level anomaly scores The pixel-level difference map was obtained by bilinear upsampling to 512×512 pixels. Weighted fusion to generate a defect confidence map The fusion formula is as follows: In the formula, The fusion weight is fixed at 0.6. The fusion result retains both pixel-level positioning accuracy and feature-level discrimination reliability, and outputs the final defect map.

[0044] Furthermore, such as Figure 4 As shown, the defect spatiotemporal correlation analysis module constructs a spatiotemporal distribution matrix from the temporally continuous defect map, and extracts periodicity, clustering, and directionality features through two-dimensional Fourier transform. The defect spatiotemporal distribution matrix is ​​denoted as... The number of rows in the matrix corresponds to the pixel coordinates in the width direction of the PVC flooring film, which is fixed at 2048 rows, corresponding to the actual width of 1024mm to 2048mm; the number of columns in the matrix corresponds to the production time step, which is fixed at 1024 columns, with each column corresponding to a flooring film with a length of 1mm; the matrix element values ​​are the defect confidence of the corresponding position, ranging from 0 to 1.

[0045] For the spatiotemporal distribution matrix Performing a two-dimensional Fourier transform converts the spatial domain distribution into a frequency domain distribution, yielding the spectrum. The formula is as follows: In the formula, For the frequency axis in the width direction, the value ranges from 0 to 2047; The length direction frequency axis has a value range of 0 to 1023; , These represent the row and column dimensions of the matrix, respectively.

[0046] Peak coordinates in the spectrum The strongest periodicity characteristic corresponding to the defect, Characterizing the periodic frequency in the width direction, The periodic frequency along the length direction is used to characterize the defect cycle length. The defect cycle length is calculated using both the production line speed and the periodic frequency along the length direction, as shown in the following formula: In the formula, This is the physical cycle length of the defect. This value directly corresponds to the circumference length of the calendering roll, traction roll, embossing roll, and cooling roll, and can accurately locate the equipment unit where the defect occurs.

[0047] The system performs three-dimensional feature quantification on defect distribution: periodicity is represented by the peak amplitude of the spectrum, ranging from 0 to 1; clustering is represented by the variance of defect confidence along the width direction, with a larger variance indicating a more concentrated defect distribution; and directionality is represented by the mean of the defect gradient direction, ranging from 0° to 180°. After normalizing these features, thermal data of defect distribution is generated. The thermal data is a 2048×1024 pixel two-dimensional matrix, with higher values ​​representing higher defect probability and severity, providing stable input for subsequent process parameter deviation tracing.

[0048] Furthermore, the process parameter deviation tracing model is a multi-input multi-output deep regression model, used to establish a precise mapping relationship between defect distribution characteristics and process parameter deviations. The model input is the defect distribution feature vector. The vector has a 32-dimensional dimension and includes quantitative indicators such as defect density, periodicity, cluster center coordinates, width offset, length interval, confidence mean, variance, and gradient direction. All features are normalized to the interval between 0 and 1. The vector formula is as follows: The model output is a vector of process parameter deviations. It includes five types of core process parameter deviations: extrusion temperature deviation, screw speed deviation, traction tension deviation, calender roll gap deviation, and cooling air temperature deviation. The vector formulas are as follows: The model employs a 4-layer fully connected network structure, with hidden layer dimensions of 64, 128, 64, and 32 respectively. The LeakyReLU activation function is used, and the output layer uses linear activation to ensure continuous numerical output. The model is trained offline using historical production data, with the training set containing no fewer than 500,000 pairs of defect feature-process deviation pairs. The Adam optimizer is used, with a fixed learning rate of 0.001, and the mean squared error loss function. The training objective is to minimize the error between the predicted and actual deviations.

[0049] The model learns to obtain a fixed nonlinear mapping relationship. During online operation, the defect distribution feature vector will be extracted in real time. Input the model and directly output the prediction deviation of the current process parameters: , Includes deviation sign, deviation value, and deviation confidence level. The system prioritizes parameters based on the absolute value of the deviation, adjusts high-impact parameters first, and generates standard industrial control instructions. The control instruction format conforms to the Modbus RTU protocol and can be directly sent to the production line PLC control system.

[0050] Furthermore, the staged PID closed-loop process control module converts the deviation output from the source-tracing model into target process parameters, and achieves smooth and precise adjustment through two-stage coarse-fine control. The formula for calculating the target process parameters is as follows: In the formula, These are the real-time acquired values ​​of the current process parameters, with an acquisition accuracy of ±0.1%. To adjust the target value; Output the deviation value for the model.

[0051] The system adopts a unified staged PID control law, the formula of which is as follows: In the formula, , representing the coarse adjustment stage and the fine adjustment stage; This represents the error between the current process parameters and the target value. The output control quantity corresponds to the actuator opening degree, ranging from 0 to 100%.

[0052] Coarse-tuning stage parameter settings: Scale coefficient Integral coefficient Differential coefficients This is used to quickly reduce errors and improve response speed. When the absolute value of the error is less than 0.5%, it automatically switches to the fine-tuning stage. Fine-tuning stage parameter settings: proportional coefficient. Integral coefficient Differential coefficients It is used to suppress overshoot, eliminate steady-state error, and ensure that parameters reach the target value stably without shock or oscillation.

[0053] The control cycle is fixed at 10ms, which is completely synchronized with the cycle of image acquisition, defect identification, and data tracing. The parameter adjustment process is written to the production database in real time for model iteration and optimization, realizing closed-loop adaptive optimization of the entire process from defect detection and root cause identification to process correction.

[0054] This implementation utilizes adaptive imaging adjustment, BRDF optical separation, feature space reconstruction, and dynamic boundary constraints to stably acquire high-definition, low-interference defect images, accurately identify and locate minute surface defects, and rapidly pinpoint the root cause of defects and infer corresponding parameter deviations based on spatiotemporal correlation analysis and process parameter deviation tracing models. Combined with staged PID closed-loop control, it achieves real-time adaptive correction of the process, automating the entire process of detection, analysis, tracing, and control. This effectively improves defect detection accuracy and production stability, ensuring consistent product quality and efficient and reliable operation of the production system.

[0055] Example 2, based on Example 1, to verify the detection accuracy of this application under real industrial conditions, selected eight existing defect detection technologies widely used in the industry for comparative testing. These technologies are: Fixed-Illu (FI), Global-Thu (GT), SVM-Def (SVM), BP-Net (BP), U-Net (UNet), Fast-RCNN (FRCNN), AE-Def (AE), and Unconstrained Anomaly Detection (UAD). This example was conducted under ideal conditions: a constant production line speed of 60 m / min, an ambient illuminance of 500 ± 50 lux, and no dust or strong reflections. In this embodiment, a standard sample of PVC floor membrane containing four typical defects—crystal points, bubbles, scratches, and pitting—was selected. The smallest defect size was 0.01 mm. The imaging quality, recognition accuracy, and real-time performance of the nine schemes were tested, and the results are shown in Table 1.

[0056] Table 1. Comparison of performance indicators of different testing schemes under standard operating conditions As shown in Table 1, this application significantly outperforms existing technologies in terms of motion blur radius (MR), defect contrast (CD), signal-to-noise ratio (SNR), and various recognition indicators.

[0057] like Figure 5As shown in the figure, there are nine independent, continuous, smooth curves, corresponding to FI, GT, SVM, BP, UNet, FRCNN, AE, UAD, and this application (This-App). The curve trends show that as the defect size decreases from 0.50 mm to 0.01 mm, the SNR curves of all schemes exhibit a downward trend. However, the curves of the prior art (FI to UAD) show a sharp inflection point drop when the size is less than 0.1 mm. Specifically, the SNR of the FI and GT curves drops below 15 dB at 0.05 mm, entering the undetectable range. In contrast, the curve of this application maintains extremely high linearity and stability throughout the entire size range. Even at a tiny defect of 0.01 mm, its SNR curve remains above 26 dB, highly consistent with the average value of 28.7 dB in Table 1. This application demonstrates excellent sensitivity and signal-to-noise ratio advantages in the extraction of tiny defect features.

[0058] To further verify the classification accuracy of the algorithm in this application from the perspective of feature space distribution, such as Figure 6 As shown, including Figure 6 (a) Figure 6 (b) Figure 6 (c) The horizontal axis represents the first principal component (PC1) of the eigenvector, and the vertical axis represents the second principal component (PC2), used to demonstrate the separability of the high-dimensional features after dimensionality reduction. The three sub-images correspond to three defect types: crystal points, bubbles, and scratches, respectively. Figure 6 As can be seen from the scatter plot of crystal point defect features in (a), the scatter clusters of traditional methods such as FI and GT are extremely diffuse, and the scatter points of normal samples and defect samples overlap in large areas, making it impossible to form a clear decision boundary; while the scatter clusters (This-App) of this application exhibit a highly compact aggregation form, and normal samples and defect samples are completely separated in space with extremely small distances within the clusters. Figure 6 (b) Scatter plot of bubble defect features. Due to the specular reflection properties of bubbles, there is significant interference. The scatter plot distribution of the prior art shows obvious tailing, indicating that feature extraction is severely affected by illumination. The scatter plot distribution of this application still maintains a regular circular or elliptical shape, and the offset of the center position is extremely small. Figure 6 (c) Scatter plot of scratch defect features: Scratches are texture-type defects and are the most difficult to distinguish. Although the scatter plots of the UAD and AE schemes have a certain degree of clustering, the edges are blurred; the scatter plot of this application shows that the defect class and the background class have a huge distance on the PC1 axis, with extremely high separation. The scatter plot intuitively shows that this application achieves a high accuracy of 97.8%; As can be seen from this embodiment, in the continuous 3-hour standard working condition stability test, the system of this application did not show obvious missed detections, false detections and parameter drift, and the overall performance fluctuation was less than 0.5%, which fully demonstrates that the application has extremely high reliability in long-term continuous online detection scenarios.

[0059] Example 3: To verify the transition from an ideal environment to a real industrial scenario, the production line speed fluctuated randomly within the range of 30~110m / min; the ambient illuminance changed abruptly between 300~800 lux; the membrane surface was covered with dust and had strong specular reflection interference; and the five core process parameters continuously drifted randomly.

[0060] Under the above-mentioned stringent conditions, the eight existing technologies in Example 2 were compared to obtain the defect cycle calculation accuracy, roller system matching success rate, process traceability deviation and closed-loop adjustment effect, as shown in Table 2. Table 2 Comparison of source tracing and control performance of different schemes under harsh working conditions As shown in Table 2, under harsh working conditions, the defect cycle calculation error (δ) of this application is only 0.17 mm, the closed-loop adjustment response time (T_resp) is only 190 ms, and the final yield is stable at 99.2%, which is comprehensively superior to the control scheme.

[0061] like Figure 7 As shown in the figure, the scatter plots representing the calculation results of nine schemes are densely distributed. It can be seen that the scatter plots of traditional schemes such as FI and GT exhibit a disordered, wide-band random distribution across the entire vertical axis range (0~3.0mm), indicating that they completely lose their ability to lock onto the defect cycle under speed fluctuations and sudden changes in illumination. While the scatter plots of the UAD and AE schemes are mainly concentrated below 1.0mm, a large number of outliers are still scattered at higher positions. In stark contrast, the scatter plots of this application are highly densely distributed in a region very close to the horizontal axis (around 0.17mm), forming a very narrow horizontal band with almost no outliers. This set of scatter plots strongly demonstrates that the spatiotemporal correlation analysis algorithm of this application still possesses extremely high robustness and computational accuracy under strong noise interference.

[0062] Furthermore, identification is performed using the aforementioned technical solutions, such as... Figure 8 As shown, Figure 8 The image on the left is the original image acquired under harsh working conditions. The image shows obvious areas of strong reflection, blurry trails caused by dust noise and speed fluctuations, and defect features are severely obscured. Figure 8 The right side shows a comparison of the processing effects of this application and eight prior art technologies, presented using grayscale confidence encoding. It is clear from the image that the confidence distribution of the regions in this application is uniform and continuous, and the defect outlines are complete and clear. In contrast, the prior art technologies all exhibit varying degrees of breakage, offset, false detections, and missed detections; some solutions even show large areas of invalid identification. This fully demonstrates that this application still possesses stable and reliable defect characterization capabilities even under extremely harsh environments.

[0063] To further verify the convergence trajectory and dynamic process of the closed-loop process control under harsh operating conditions, such as Figure 9 The phase trajectory curves shown exhibit severe oscillations and slow convergence when using traditional schemes such as FI and GT. Even after 120 seconds, the curves are not fully stable, and the residuals are large. This corresponds perfectly to the steady-state error (Ess) of up to 4.82% for the FI scheme in Table 2. When using schemes such as UNet and FRCNN, although the convergence speed is improved, there is obvious overshoot (the initial curve significantly crosses the zero axis, forming peaks / troughs) and continuous small oscillations. In contrast, the curves of this application (This-App) perform exceptionally well: at the moment of interference, the curve responds rapidly and smoothly returns to near zero. The entire process is free of overshoot and oscillations, and it stabilizes completely in a very short time (approximately corresponding to the shortest response time Tresp of 190ms in Table 2), with the steady-state error approaching zero. This curve intuitively demonstrates the significant advantages of the strategy of this application in terms of dynamic response speed and control stability.

[0064] This embodiment operated continuously for 12 hours under harsh combined working conditions. The system did not experience data loss, parameter drift, control interruption, or identification failure. The overall defect detection rate remained above 99.5%, and the final yield rate was consistently above 99.2%. This fully demonstrates that the application has high robustness, reliability, and feasibility in real-world harsh industrial environments.

[0065] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A closed-loop optimization method for production processes based on the detection of surface defects in PVC flooring films, characterized in that, include: Image data of PVC floor membrane surface is collected. Based on the real-time running speed and surface texture type of PVC floor membrane, the incident angle of the light source and the exposure parameters of the camera are adaptively adjusted by a dynamic structured light source and a line scan camera to obtain image data of minute surface defects under different working conditions. The collected surface image data is preprocessed. An ambient light interference suppression model is established by using an adaptive image enhancement algorithm based on the physical imaging model and combining it with the optical reflection characteristics of PVC material. By separating the specular reflection component and the diffuse reflection component, the texture and morphology features of the defect area are obtained, and a set of feature vectors to be tested is constructed. The set of feature vectors to be tested is input into a pre-trained defect recognition model. The defect recognition model adopts an anomaly detection architecture based on feature space reconstruction. By introducing a dynamic boundary constraint mechanism in the normal sample feature space, the pixel-level reconstruction error between the feature vectors to be tested and the feature distribution of normal samples is calculated, and a defect map is generated. Based on the defect map, the distribution of defects in the width and length directions of the PVC floor film is extracted using a spatiotemporal correlation analysis algorithm. By analyzing the periodicity, clustering, and directionality of the defect distribution, the potential root causes of the defects are identified, and thermal data of the defect distribution is generated. The defect distribution thermal data is correlated and mapped with the production process parameter space to construct a process parameter deviation tracing model. By analyzing the correlation between defect distribution characteristics and historical data of process variables, the type and degree of process parameter deviation that caused the defect are inferred, and control instructions are generated. The control commands are sent to the production line control system in real time to adjust the production process parameters of the PVC floor film in stages, thereby achieving closed-loop adaptive optimization of the production process. The process of adaptively adjusting the incident angle of the light source and the camera exposure parameters using a dynamic structured light source and a line scan camera based on the real-time operating speed and surface texture type of the PVC floor membrane includes: Real-time acquisition of production line operating speed and preset PVC floor membrane surface texture type ; According to the running speed Calculate the camera's exposure time This makes the motion blur radius Less than the preset fuzzy threshold ,in , The magnification of the optical system; According to the surface texture type Query the preset illumination angle mapping table to determine the target incident angle of the structured light source. The illumination angle mapping table stores the correspondence between different texture types and the optimal incident angle, and the optimal incident angle is used to enhance the defect contrast under a specific texture. According to the running speed Calculate the camera's exposure time It also includes: Obtain the camera's signal-to-noise ratio With exposure time nonlinear response function ,in Light source intensity; In satisfying Determine the optimal exposure time under the constraints. : ,in, and These are weighting coefficients, used to balance the trade-off between image signal-to-noise ratio and motion blur.

2. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 1, characterized in that, The adaptive image enhancement algorithm based on the physical imaging model, combined with the optical reflectivity of PVC material, establishes an ambient light interference suppression model, including: Construct a physical imaging model based on the bidirectional reflectance distribution function (BRDF) to measure the intensity of the acquired images. Represented as ambient light component Specular reflection component and diffuse reflection component The linear superposition is given by the formula: ; Multiple frames of images are acquired using a polarization camera or multi-angle light source, and the specular reflection component is separated using the polarization difference method or photometric stereo method. and diffuse reflection component ; Ambient light component Background modeling and subtraction were performed to obtain a defect feature image free from ambient light interference. The formula is: ,in This represents the average grayscale value of the background area.

3. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 2, characterized in that, The process involves acquiring multiple frames of images using a polarization camera or multi-angle light source, and then separating the specular reflection component using the polarization difference method. and diffuse reflection component ,include: Controlling the polarization camera in two orthogonal polarization directions and Images were acquired separately below and ; Calculate the degree of polarization based on Fresnel reflection principle. and polarization angle Among them, the degree of polarization Used to characterize the intensity of specular reflection: Based on the degree of polarization Construct a specular reflection mask Regions with polarization degrees higher than a preset threshold are identified as specular reflection regions, thereby separating them. .

4. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 1, characterized in that, The defect identification model adopts an anomaly detection architecture based on feature space reconstruction, which introduces a dynamic boundary constraint mechanism into the feature space of normal samples, including: Construct a feature extraction network based on an autoencoder and a decoding network Input image Mapping to feature space Then reconstructed into an image ; In feature space In the middle, define a dynamic boundary sphere. ,in The center of the normal sample features, The boundary radius; During the training phase, minimize the reconstruction error of normal samples. At the same time, constrain the characteristics of normal samples Falling inside the boundary sphere means minimizing the boundary constraint loss. ; During the inference phase, the features of the test sample are calculated. To the center distance ,like If so, it is judged as abnormal.

5. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 4, characterized in that, The calculation of the pixel-level reconstruction error between the feature to be tested and the feature distribution of normal samples, and the generation of a defect map, includes: Computational reconstruction of the image Compared with the original input image Pixel-level difference map ; Calculate the features to be measured Anomaly score in feature space ,in The standard deviation of the characteristics of normal samples; pixel-level difference map With feature-level anomaly scores Weighted fusion is performed to generate a defect confidence map. The formula is: ,in, To integrate weights, This is an operation that upsamples the feature map size to the original image size.

6. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 1, characterized in that, The method of extracting the distribution of defects in the width and length directions of the PVC flooring membrane using a spatiotemporal correlation analysis algorithm includes: By arranging continuous defect maps in a time series, a spatiotemporal distribution matrix of defects is constructed. The rows of the matrix correspond to the width coordinates of the production line, and the columns correspond to the time steps. For the spatiotemporal distribution matrix of the defects Perform a two-dimensional Fourier transform to obtain the spectrum. ; Analyze the spectrum peak position ,in The corresponding periodic frequency of the defect along its length. The periodic frequency of the corresponding defect in the width direction; According to the periodic frequency Calculate the period length of the defect , To determine the operating speed of the production line, identify periodic defects associated with the traction roller or embossing roller.

7. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 1, characterized in that, The constructed process parameter deviation tracing model, by analyzing the correlation between defect distribution characteristics and historical data of process variables, inversely deduces the type and extent of process parameter deviations that lead to defects, including: Construct a multi-input, multi-output regression model, with the input being the defect distribution feature vector. The output is a vector of process parameter deviations. ; The regression model is trained using historical production data to learn the mapping relationship from defect characteristics to process deviations; During the online detection phase, the defect distribution feature vector is extracted in real time. Input the trained model to predict the current process parameter deviation. .

8. The closed-loop optimization method for production process based on PVC flooring film surface defect detection according to claim 7, characterized in that, The process of sending control commands to the production line control system in real time and adjusting the production process parameters of the PVC flooring film in stages includes: Predicted process parameter deviations This is converted into a control command, which includes adjusting the target value. ; A phased PID control strategy is adopted, dividing the adjustment process into two stages: coarse adjustment and fine adjustment. In the coarse adjustment phase, a larger scaling factor is used. Rapidly approach the target value; when the error between the process parameters and the target value is less than the preset threshold, switch to the fine-tuning stage and use a smaller scaling factor. and integral coefficient To eliminate steady-state error, the control law is: ,in, This represents the error in the process parameters at the current moment. This is the current adjustment phase.

Images