A method and system for identifying surface defects of a corrugation roll

Abstract

The present application relates to the technical fields of intelligent manufacturing and machine vision detection, and discloses a corrugated roller surface defect recognition method and system. The method comprises: acquiring multiple two-dimensional texture images of the corrugated roller surface through multiple-angle light sources; inverting the surface normal vector based on the light intensity distribution and reconstructing the three-dimensional micro-topography point cloud; converting the point cloud into a topological graph structure; and using a graph neural network to recognize abnormal topological paths to output a defect result. The system comprises an image acquisition module, a three-dimensional reconstruction module, a topological construction module and a defect recognition module. By using the above technical solution, the present application can realize high-precision three-dimensional reconstruction and topological path analysis of micro-defects, and improve the detection robustness and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and machine vision inspection technology, specifically relating to a method and system for identifying surface defects on corrugated rolls. Background Technology

[0002] With the automation and intelligent upgrading of the packaging industry, the operational precision and efficiency of corrugated cardboard production lines have become core indicators for measuring the competitiveness of packaging companies. As a key component in corrugated cardboard forming equipment, the surface quality of the corrugated rollers determines the forming effect, physical strength, and subsequent printing flatness of the cardboard. In the long-term, high-frequency industrial production environment, defects such as wear, scratches, and micro-cracks on the corrugated rollers are unavoidable. High-precision real-time detection and monitoring of the corrugated roller surface is an important link in ensuring the stable operation of the production line and improving product quality.

[0003] Corrugated roll surface defect identification technology aims to monitor the surface of corrugated rolls in real time under high-speed rotation through high-precision imaging and intelligent recognition algorithms. This technology utilizes specific light sources and sensing devices to capture the geometric features and texture changes of the surface, and combines this with computer vision technology to accurately locate and classify various micro-defects. Its goal is to replace traditional manual visual inspection with non-contact measurement methods, thereby achieving closed-loop quality control in the production process.

[0004] Existing surface defect identification technologies for corrugated rolls suffer from several limitations: First, the imaging mode is too simplistic. Traditional two-dimensional optical inspection techniques are prone to localized overexposure due to specular reflection when dealing with highly reflective metal surfaces that have undergone quenching or chrome plating. Second, the uniformity of the metal's natural color results in low image contrast, making it difficult to detect shallow scratches or minute indentations. Third, feature recognition capabilities are insufficient. Corrugated roll surfaces exhibit regular, undulating tooth-like textures. Conventional algorithms often fail to distinguish normal tooth edges from abnormal defects when processing images with such textured backgrounds, leading to a high false alarm rate. Fourth, three-dimensional morphology reconstruction is lacking. For hidden cracks extending along the tooth root, the lack of depth information makes it difficult to accurately depict their topological path using only two-dimensional textures, easily resulting in missed detections. Fifth, the detection system lacks the ability to model complex spatial structures, making it difficult to accurately reconstruct minute defects in highly disruptive industrial environments. In view of the shortcomings of traditional corrugated roll surface defect identification technology, such as overly simple imaging modes, insufficient feature recognition capabilities, lack of three-dimensional morphology reconstruction, and weak spatial modeling capabilities, the corrugated roll surface defect identification method and system proposed in this invention are of great importance. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying surface defects on corrugated rolls, which can solve the problems mentioned in the background art. Addressing the high reflectivity of corrugated roll surfaces, the low contrast caused by the uniform metallic color, and the interference in defect identification due to regular tooth-shaped textures, this invention achieves high-precision three-dimensional reconstruction and topology path analysis of micro-defects on the corrugated roll surface by deeply fusing photometric stereo vision technology and graph neural networks, thereby reducing the false detection rate and the missed detection rate.

[0006] To achieve the above objectives, this invention proposes a method for identifying surface defects on corrugated rolls, comprising the following steps: S1, acquiring multiple frames of two-dimensional texture images of the corrugated roll surface under illumination from multiple predetermined angle light sources through a preset illumination sequence; S2, performing surface normal vector inversion based on the reflected light intensity distribution information in the multiple frames of two-dimensional texture images, and reconstructing the three-dimensional micro-topography point cloud of the corrugated roll surface accordingly; S3, transforming the three-dimensional micro-topography point cloud into a topological graph structure with spatial connectivity, where sampling points in the point cloud serve as graph nodes, and geometric neighborhood relationships between sampling points serve as edges of the graph; S4, performing feature extraction and message passing on the topological graph structure through a graph neural network, identifying and extracting abnormal topological paths representing defects, and outputting the defect identification result.

[0007] Preferably, step S2 specifically includes the following steps: S21, constructing a surface reflection model based on a predetermined reflection law, and establishing a functional mapping relationship between the reflected light intensity of each pixel and the surface normal vector, the light source direction vector, and the surface reflectivity; S22, solving the light intensity equations of each pixel under multiple predetermined angle light sources using the least squares method, and separating the surface reflectivity information and the surface normal vector information; S23, performing spatial integration on the surface normal vector information to obtain the relative height field of the corrugated roller surface, and generating a three-dimensional micro-morphological point cloud in combination with the system calibration parameters.

[0008] Preferably, the predetermined reflection law mentioned in step S21 is a modified Lambert's reflection law, which introduces a specular reflection correction term for highly reflective metal surfaces based on the basic reflection model. The specular reflection correction term controls the weight of the specular component through a preset attenuation coefficient to counteract the strong reflective interference caused by quenching or chrome plating on the corrugated roll surface.

[0009] Preferably, the process of solving the surface normal vector information in step S22 includes: performing brightness normalization processing on the acquired multi-frame two-dimensional texture images to eliminate errors caused by fluctuations in light source intensity; constructing a linear equation system with the normalized light intensity value and its corresponding light source direction vector; and optimizing the linear equation system by using singular value decomposition to obtain the normal vector component of each pixel in three-dimensional space.

[0010] Preferably, step S23, which involves spatial integration of the surface normal vector information, specifically includes: employing a normal integration algorithm to transform the normal vector field into a continuous height map by minimizing the difference function between the surface gradient and the measured normal vector. During the integration process, a gradient operator is introduced to smooth the tooth edges of the corrugated roll, thereby avoiding numerical integration divergence in areas with high curvature, such as the tooth tip and tooth root.

[0011] Preferably, step S3 specifically includes the following steps: S31, downsampling the three-dimensional micro-topography point cloud, retaining feature points that reflect surface details, and constructing a set of nearest neighbor nodes for each feature point according to a preset search radius; S32, calculating the geometric feature differences between nearest neighbor nodes, wherein the geometric feature differences include height gradient deviation, normal angle, and local curvature change; S33, assigning initial weights to the edges in the topological graph structure based on the geometric feature differences, with regions of large geometric abrupt changes being given high initial weights to initially identify potential crack or scratch areas.

[0012] Preferably, step S4 specifically includes the following steps: S41, using graph convolutional layers to aggregate node features in the topological graph structure, each graph convolutional operation updates the feature representation of the current node by aggregating the geometric information of adjacent nodes; S42, by stacking multiple graph convolutions, capturing topological features with long-distance dependency characteristics, which are used to characterize the physical path of cracks extending along the root of the corrugated roll teeth; S43, inputting the updated node features into a fully connected layer and a classifier to determine the probability that each node belongs to normal texture, crack defect, scratch defect or wear defect.

[0013] Preferably, the node feature aggregation process in step S41 employs an attention mechanism. This attention mechanism is configured to automatically learn the association weights between adjacent nodes, assigning larger weights to adjacent nodes that conform to crack topological features, and smaller weights to adjacent nodes that conform to normal tooth-like patterns, thereby enhancing defect signals and suppressing normal background texture signals in the graph structure.

[0014] Preferably, capturing long-distance dependency characteristics in step S42 includes: introducing residual connections and skip connections in the graph neural network to ensure that the original three-dimensional topographic details can be preserved in the deep network; and using global pooling operations to extract global statistical features of the entire topology graph to help improve the recognition accuracy of large-area wear defects.

[0015] Preferably, after outputting the defect identification results, the method further includes a step of quantitatively evaluating the identified defects. The quantitative evaluation includes calculating the depth, width, and extension length of the defect. For crack-type defects, the crack extension trajectory is determined by searching the longest continuous abnormal path in the topology graph.

[0016] A corrugated roll surface defect recognition system, used to implement the above-mentioned method, includes: an image acquisition module configured to acquire multiple frames of two-dimensional texture images of the corrugated roll surface in cooperation with multiple predetermined angle light sources; a three-dimensional reconstruction module, whose input end is connected to the image acquisition module, used to perform surface normal vector inversion and height field integration based on photometric stereo vision, and output a three-dimensional micro-topography point cloud; a topology construction module, whose input end is connected to the three-dimensional reconstruction module, used to convert the point cloud data into a topological graph structure and calculate the initial weights of the edges; and a defect recognition module, including a pre-trained graph neural network, used to receive the topological graph structure and output the classification and localization results of defects.

[0017] Preferably, the image acquisition module includes multiple light sources arranged in a ring and a high-speed industrial camera located at the center. The activation time of the multiple light sources and the exposure time of the high-speed industrial camera are precisely triggered by a synchronization controller to ensure that the sampling area of ​​the same surface of the corrugated roll can be covered by light sources at different angles during the rotation of the corrugated roll.

[0018] Preferably, the 3D reconstruction module incorporates a normal vector correction operator, which compensates for secondary reflected light generated on the metal surface. By analyzing the reflection path of light on the side of the corrugated roll teeth, interfering light intensity is eliminated, improving the reconstruction accuracy in the shadow area at the tooth root.

[0019] Preferably, the graph neural network in the defect recognition module is trained using pre-collected corrugated roll samples containing known defect types. During training, the cross-entropy loss function is used to measure the difference between the predicted and true categories, and the network parameters are updated using a backpropagation algorithm until the model's recognition accuracy reaches a preset threshold.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention upgrades traditional two-dimensional texture detection to three-dimensional microscopic morphology reconstruction based on optical physics inversion by introducing photometric stereo vision technology.

[0021] 2. Irradiation from multiple angles can effectively break the interference of specular reflection under a single viewpoint. The normal vector field obtained by inverting the light intensity distribution is extremely sensitive to the minute geometric deformation of the surface. This enables the present invention to identify shallow scratches and micro-indentations that are almost invisible in traditional two-dimensional images.

[0022] 3. By transforming 3D point clouds into a topological graph structure and utilizing graph neural networks for analysis, this invention changes the paradigm of feature extraction. The node connectivity characteristics of graph neural networks can perfectly fit the physical extension features of defects, enabling the algorithm to automatically distinguish between regular, normal toothed edges and abnormal cracks with random topological paths. This invention achieves an organic unity of physical modeling and deep learning, improving the robustness of defect detection in complex industrial contexts.

[0023] 4. The photometric stereoscopic vision technology employed in this invention does not rely on expensive structured light equipment. It achieves sub-micron level depth resolution simply through a logical combination of ordinary light sources, reducing the hardware cost of the system. By using a specular reflection correction term in the reflection model, the overexposure problem on the highly reflective chrome-plated corrugated roller surface is solved, ensuring the stability of image quality. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of defect recognition based on the fusion of photometric stereo vision and graph neural network in this invention; Figure 3 This is a flowchart illustrating the logical process of reconstructing the three-dimensional microstructure of the corrugated roll surface based on a multi-source reflection model in this invention. Figure 4 This is a feature extraction logic framework diagram for the construction of a 3D point cloud topology map and the identification of abnormal paths in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the corrugated roller surface defect identification system in this invention. Detailed Implementation

[0025] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0026] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 This embodiment provides a method for identifying surface defects on corrugated rolls, primarily applied in corrugated cardboard production lines for online monitoring of the surface quality of corrugated rolls, a core component. Corrugated rolls, as metal components with complex geometric topologies, typically undergo quenching or chrome plating, resulting in gloss and reflectivity, and their surfaces are covered with regular tooth-shaped ripples. This embodiment achieves the identification of minute defects in this complex environment by integrating photometric stereo vision technology with graph neural networks.

[0027] Step S1 involves acquiring multiple frames of two-dimensional texture images of the corrugated roll surface under illumination from multiple predetermined angle light sources using a preset illumination sequence. In practice, the image acquisition system is installed radially outside the corrugated roll. The illumination sequence consists of an array of eight LEDs evenly distributed in a ring, each with a different incident angle relative to the normal direction of the corrugated roll surface. As the corrugated roll rotates at a predetermined angular velocity, the synchronous controller precisely triggers the eight light sources to flash sequentially at high frequency based on the pulse signal from the rotary encoder. During each flash, a high-speed industrial camera located at the center of the ring of light sources performs an exposure. Due to the varying angles of the light sources, the same sampling area on the corrugated roll surface exhibits drastically different brightness and shadow characteristics under different illumination conditions. The acquired multiple frames of two-dimensional texture images are cached in a high-performance image processing unit, with each frame having a pixel resolution of no less than 4096×3072 to ensure the capture of sub-millimeter-level surface details.

[0028] Step S2 involves inverting the surface normal vector based on the reflected light intensity distribution information in multiple frames of two-dimensional texture images, and reconstructing the three-dimensional microscopic topography point cloud of the corrugated roll surface accordingly. This step is crucial for transitioning from two-dimensional vision to three-dimensional physical modeling, and its core lies in utilizing the principles of photometric stereo vision. Specific sub-steps include: S21. Construct a surface reflection model based on a predetermined reflection law, establishing a functional mapping relationship between the reflected light intensity of each pixel and the surface normal vector, the light source direction vector, and the surface reflectivity. In this embodiment, an improved Lambert reflection law is used to describe the optical properties of the metal surface. The conventional Lambert model only considers the diffuse reflection component, while this invention introduces a specular reflection correction term for highly reflective metal surfaces. This correction term controls the weight of the specular component through a preset attenuation coefficient. The observed light intensity at a certain point on the surface is defined as the product of the reflectivity, the normal vector, and the inner product of the light source direction vector, with an additional correction term constrained by the specular reflection coefficient and the high-brightness direction to counteract the strong reflective interference caused by the chrome plating treatment of the corrugated roller surface.

[0029] S22. The light intensity equations for each pixel under multiple predetermined angle light sources are solved using the least squares method to separate surface reflectivity information and surface normal vector information. During processing, the brightness of the acquired multi-frame two-dimensional texture images is normalized to eliminate systematic errors caused by inconsistent luminous efficiency of each light source or fluctuations in driving voltage. For each pixel location, a linear equation system is constructed by combining the corresponding light intensity values ​​from 8 frames of images with the known direction vectors of 8 light sources. This linear equation system is optimized using singular value decomposition to calculate the unit normal vector pointing outwards at that point while minimizing reconstruction error. The three components of the normal vector represent the spatial orientation of the tiny surface region in the horizontal, vertical, and depth directions, respectively.

[0030] S23. Perform spatial integration on the surface normal vector information to obtain the relative height field of the corrugated roll surface, and generate a three-dimensional microscopic topographic point cloud by combining it with system calibration parameters. This embodiment uses a variational normal integration algorithm. This algorithm transforms the discontinuous normal vector field into a continuous height map by minimizing the difference function between the surface gradient and the measured normal vector. During the integration iteration process, a specific gradient operator is introduced to smooth the tooth edges of the corrugated roll. The corrugated roll has curvature changes at the tooth tip and tooth root positions, and conventional integration is prone to numerical divergence, while the gradient operator can ensure that stable depth information can still be obtained in these areas. The generated three-dimensional point cloud data not only contains the surface texture, but also accurately depicts the microscopic geometric undulations.

[0031] Then, step S3 is performed to transform the 3D microscopic point cloud into a topological graph structure with spatial connectivity. Unlike traditional pixel-based mesh processing methods, the topological graph structure can more naturally express the physical connectivity characteristics of defects on the metal surface.

[0032] S31. The 3D microscopic topographic point cloud is downsampled to retain feature points reflecting surface details, and a set of nearest neighbor nodes for each feature point is constructed according to a preset search radius. The search radius is typically set to 0.5 mm to 1.5 mm. This process removes redundant smooth region data and preserves abrupt geometric edges.

[0033] S32. Calculate the geometric feature differences between neighboring nodes. These differences include not only the deviation in height gradient, but also the angle between the normal vectors of adjacent points and the local curvature changes. For example, on a flat corrugated roll surface, the angle between the normal vectors of adjacent nodes is close to 0 degrees; however, once a crack or scratch appears, the normal vectors of adjacent nodes will change drastically, and the angle will increase.

[0034] S33. Assign initial weights to edges in the topological graph structure based on differences in geometric features. Regions with significant geometric abrupt changes (such as edges with normal angles exceeding a preset threshold) are assigned higher initial weights. These high-weighted edges initially identify potential crack or scratch regions, providing important prior guidance for subsequent deep learning recognition.

[0035] Finally, step S4 is executed: feature extraction and message passing are performed on the topology graph structure through graph neural network to identify and extract abnormal topology paths representing defects, and the defect identification results are output.

[0036] S41. Graph convolutional layers are used to aggregate node features in the topological graph structure. Each graph convolutional operation updates the feature representation of the current node by aggregating the geometric information of neighboring nodes. This embodiment employs a multi-head attention mechanism, which can automatically learn the association weights between neighboring nodes. For neighboring nodes that conform to the physical laws of cracks (such as abnormal points continuously distributed along a specific direction), larger attention weights are assigned; while for nodes that conform to the periodic laws of normal tooth profiles, smaller weights are assigned. This approach enhances the ability to capture defect signals and suppresses background noise.

[0037] S42. By stacking multiple layers of graph convolution, topological features with long-range dependencies are captured. This is particularly important for characterizing long, narrow cracks extending along the root of corrugated roll teeth. By introducing residual connections and skip connections, it is ensured that the deep network can still retain the original three-dimensional morphological details when performing high-level semantic abstraction. Global pooling is used to extract global statistical features of the entire topological graph, assisting the system in determining whether there is large-area uniform wear.

[0038] S43. Input the updated node features into the fully connected layer and classifier to calculate the probability distribution of each node belonging to normal, cracked, scratched, or worn. By performing cluster analysis on the probability field, determine the specific boundaries and types of defects.

[0039] After outputting the defect identification results, this embodiment also includes a step of quantitatively evaluating the defects. The system automatically calculates the average depth, maximum width, and physical extension length of the defects. For crack-type defects, the growth trajectory is accurately located by searching the longest continuous anomaly weight path in the topology graph.

[0040] This embodiment also provides a supporting corrugated roll surface defect recognition system, whose hardware components include: an image acquisition module, consisting of a ring light source array located above the corrugated roll and a high-speed industrial camera at the center; a 3D reconstruction module, built into a high-performance industrial computer, executing normal vector inversion logic; a topology construction module, responsible for converting the reconstructed point cloud into a graph structure in real time; and a defect recognition module, running a pre-trained graph neural network model. This system can be linked with the PLC control system of the corrugated paper production line, and once a fatal defect is detected, it immediately issues an alarm signal and marks the defect location.

[0041] Example 2: Based on Example 1, this example further optimizes the accuracy of 3D reconstruction and the robustness of the algorithm for corrugated rolls that have been used for a long time and have secondary reflection interference on their surface.

[0042] In the deep tooth structure of corrugated rolls, light often undergoes multiple reflections between adjacent tooth surfaces, a phenomenon known as secondary reflection. This results in the light intensity value captured by the industrial camera not only including reflected light from the light source but also stray light from adjacent tooth surfaces. To address this issue, this embodiment introduces a dedicated normal vector correction operator in step S22.

[0043] The normal vector correction operator works based on the geometric analysis of the reflection path. The 3D reconstruction module predicts the regions where secondary reflections may occur at specific lighting angles using a pre-established standard geometric model of the corrugated roll. Before performing normal vector inversion, the algorithm calculates the mutual reflection components of this region and subtracts this interference term from the observed light intensity. Through this iterative correction, the reconstruction accuracy of the system in the shadow region at the root of the corrugated roll tooth is improved.

[0044] This embodiment refines the graph neural network training process described in Embodiment 1. To enable the model to adapt to corrugated rolls of different models and tooth pitches, a transfer learning strategy is introduced during training. Pre-training is performed using a large-scale publicly available 3D shape dataset to acquire basic geometric feature extraction capabilities. Fine-tuning is then performed using a specific corrugated roll dataset containing 3000 known defect samples collected from industrial sites. During fine-tuning, a weighted combination of cross-entropy loss and structural similarity loss is employed, ensuring the network focuses not only on classification accuracy but also on the accuracy of reconstructing the geometric shape of the defects.

[0045] In terms of hardware configuration, the image acquisition module in this embodiment uses a high-speed CMOS camera with a global shutter, increasing the frame rate to 500 frames per second, which can be used to perform real-time sampling in conjunction with the linear speed of the corrugated roller at 300 meters per minute. The light source uses a blue light source with a wavelength of 450 nanometers, because short-wavelength light scatters less on the metal surface, which is more conducive to extracting extremely fine surface scratch features.

[0046] Example 3: This example focuses on how the present invention achieves simultaneous identification of large-area wear and micro-point defects on the surface of corrugated rolls in a multi-algorithm parallel processing environment.

[0047] In the downsampling process of step S31, this embodiment employs a multi-scale sampling strategy. The system maintains two-scale topology maps: one is a detail map that retains high-density nodes, used to identify tiny pits and short scratches; the other is an extremely simplified macroscopic map, used to characterize large-scale geometric deviations of the surface.

[0048] During the message passing process in step S42, features from the macroscopic image are periodically fed back to the detail image. This cross-scale information fusion enables the system to identify complex composite defects, such as micro-stress cracks (detail features) that are more likely to occur in areas thinned due to wear (macroscopic features).

[0049] To further enhance the system's environmental adaptability, this embodiment introduces a dynamic confidence threshold in the classification decision stage of S43. This threshold is automatically adjusted based on the background noise level of the current ambient light. When there is strong external stray light interference in the detection environment, the system automatically raises the probability threshold for determining defects, thereby controlling the false alarm rate.

[0050] At the system implementation level, the 3D reconstruction module and defect identification module in this embodiment adopt a hardware-accelerated parallel computing architecture. The generation of 3D point clouds and the inference of graph neural networks are executed synchronously in different stream processor clusters. By adopting this pipelined data processing mode, the latency from image acquisition to the final output of the defect assessment report is controlled within 50 milliseconds, which fully meets the requirements of real-time closed-loop control of the production line.

[0051] Example 4: This example details the installation and commissioning process of the present invention in an actual production line environment and its specific compensation mechanism.

[0052] Because corrugated cardboard production lines inevitably experience mechanical vibrations during high-speed operation, these vibrations can cause spatial distortion during image acquisition. To compensate for this displacement deviation, this embodiment incorporates a laser displacement compensation device into the image acquisition module. This device can sense the distance fluctuations between the corrugated roll surface and the industrial camera in real time.

[0053] During the height field integration in step S23, the system corrects the initial integration value in real time based on submicron-level vibration data provided by the laser displacement compensation device. This dynamic compensation mechanism ensures that the generated three-dimensional microscopic point cloud maintains spatial consistency even under strong vibration conditions, avoiding interference from pseudo-defects caused by mechanical vibration.

[0054] To address potential interference from residual oil or paper dust on the corrugated roll surface, this embodiment adds a texture feature consistency verification step in step S32. This step analyzes the color moment characteristics of local areas to determine whether feature jumps are caused by geometric defects or surface coverings. If determined to be a covering, the corresponding node is assigned a very low edge weight in the topology graph, avoiding false detections caused by foreign object contamination.

[0055] On the software interface, the system provides operators with a 3D visualization view. Identified cracks are displayed as prominent red topological paths on the reconstructed 3D model of the corrugated roll. The quantitative evaluation results are synchronized in real time to the factory's quality management database, providing a scientific basis for predicting the replacement cycle of the corrugated roll and for preventative maintenance.

[0056] Example 5: This example further explores the universality optimization of the present invention for corrugated rolls made of different materials. Besides the common chrome-plated corrugated rolls, tungsten carbide-coated corrugated rolls are also widely used in industry. Tungsten carbide surfaces have an extremely fine granular texture, and their reflectivity differs from that of smooth chrome-plated surfaces.

[0057] To address this issue, this embodiment introduces an adaptive switching mechanism for the reflection model in step S21. The system automatically determines the surface material of the corrugated roll to be inspected by analyzing the grayscale histogram distribution of the first frame image. For tungsten carbide surfaces, the system switches to a reflection descriptor based on the Eulenauer model, which can better handle diffuse reflection and lateral scattering issues on rough surfaces.

[0058] In the graph convolution operation of step S41, to further enhance the contrast of defect edges, this embodiment introduces the inverse operation of the Laplacian smoothing operator. By artificially enhancing the difference in features between adjacent nodes during message passing, even extremely shallow scratches (depth less than 5 micrometers) hidden in the background texture can form clear energy ridges in the topology graph.

[0059] Regarding system maintenance, this embodiment provides an automatic calibration function. Every 24 hours, the system automatically controls the camera to image the built-in standard step block, and automatically updates the optical scale parameters in the system by comparing the measured height with the standard height.

[0060] Example 6: This example details how the present invention utilizes the global statistical properties of graph neural networks to predict the remaining life of corrugated rolls.

[0061] Following the output of step S4, the system continuously collects all detected defect distribution patterns over a period of time (e.g., a work week). The defect identification module utilizes a long short-term memory network based on spatiotemporal graph convolution to model the evolution trend of defects.

[0062] If the system observes that the crack topology path in a certain tooth root region shows an expanding trend over the past 48 hours, or if the distribution density of point defects exceeds a preset critical level, the system will not only report the current defect status but also output a life decay coefficient through the fully connected layer. This coefficient, combined with the current total paper feed of the corrugated roll, can accurately predict the specific number of hours remaining before the corrugated roll fails.

[0063] This proactive diagnostic capability enables packaging companies to shift from traditional "post-failure maintenance" to "condition-based precision maintenance," reducing unplanned downtime losses caused by sudden tooth breakage of corrugated rollers.

[0064] Example 7: This example describes the application of the present invention in a large-scale distributed inspection architecture. In large packaging groups, multiple production lines often operate simultaneously. In this example, the defect identification module is deployed on a cloud server, while the image acquisition and 3D reconstruction module is deployed at the edge of the production line.

[0065] The topology graph generated in step S3, due to downsampling, has a data volume of only about 5% of the original image data. This transforms the originally cumbersome image data transmission into lightweight graph structure transmission, alleviating network bandwidth pressure.

[0066] After receiving the topology maps from each production line, the cloud server uses massively parallel computing resources to perform deep inference and transmits the identification results back to the production site in real time. This architecture not only reduces the hardware investment cost of a single production line, but also enables the cloud model to continuously evolve from diverse defect data from different factory areas, achieving a swarm learning effect for the algorithm.

[0067] In terms of data security, the transmitted topology map data has undergone asymmetric encryption. Since the graph structure is essentially an abstract geometric connection and does not contain any visible light visual information, it also protects the company's production process parameters and technical secrets.

[0068] Example 8: This example addresses the thermal deformation problem of corrugated rolls under specific production environments and proposes a morphology restoration and correction scheme. In high-temperature production environments, corrugated rolls experience slight thermal expansion, causing their nominal tooth profile parameters to drift.

[0069] In step S23, the system introduces a temperature compensation coefficient. This coefficient is obtained through an infrared temperature sensor mounted on the corrugated roll bearing housing. The 3D reconstruction module performs a linear correction to the length reference during the height field integration process based on real-time temperature data.

[0070] This optimization ensures that the defect depth and width calculated by the system are based on the physical dimensions of the corrugated roll under standard room temperature conditions, avoiding quantitative assessment errors caused by thermal expansion and contraction. During a 12-hour continuous high-temperature production test, this compensation mechanism reduced dimensional measurement errors.

[0071] To address the air turbulence interference caused by high-speed rotation, this embodiment adds a constant-pressure air curtain to the front of the lens of the image acquisition module. This air curtain, through laminar gas blowing, forms a stable optical medium layer between the lens and the corrugated roll, isolating the refraction effects of paper dust, oil mist, and hot airflow on the light path, thus ensuring the edge sharpness of multi-frame two-dimensional texture images.

[0072] Example 9: This example details the underlying logic of the attention mechanism in graph neural networks in suppressing normal texture backgrounds. The toothed texture of corrugated rolls has extremely strong periodicity. This periodic signal is easily superimposed on the spectrum generated by linear defects in traditional frequency domain filtering algorithms, making it difficult to separate.

[0073] In step S41 of this embodiment, the attention mechanism is configured as a spatially selective filter. During the training phase, the network learns a topological template of the normal tooth profile of a corrugated roll. During inference, when a node connection conforming to the periodic characteristics of this template is detected, the attention head produces a low response value.

[0074] Conversely, when a topological connection that breaks this geometric periodicity occurs (such as an oblique crack spanning the tooth flank), the attention head generates an extremely high excitation signal. This filtering method based on topological logic essentially utilizes the inconsistency in the spatial evolution of defects and the background to solve the problem of defect detection under strong background texture interference.

[0075] By stacking multiple attention features, the system can identify subpixel-level texture anomalies that are almost indistinguishable to the naked eye, providing the most detailed metadata for high-precision repair of corrugated rolls.

[0076] Example 10: This example illustrates the method of achieving detection consistency through algorithmic adaptive adjustment under different lighting conditions according to the present invention.

[0077] Because ambient light in the workshop (such as sunlight, fluorescent lights, etc.) changes over time, it can cause environmental noise interference to the two-dimensional texture image acquired in step S1. In this embodiment, the system sets a reference white balance area in the image acquisition module.

[0078] During each frame of image acquisition, the algorithm reads the average light intensity of the reference area and adjusts the global ambient light constant term in the reflection model of step S21 accordingly. Through this real-time feedback correction, the system ensures that the surface normal vector field obtained by the inversion remains highly consistent regardless of whether it is a day shift or a night shift.

[0079] During the topology construction phase, the weight allocation logic in step S33 also incorporates adaptive signal-to-noise ratio correction. For sampling regions with low signal-to-noise ratio, the system automatically increases the integration step size for geometric feature difference determination, using spatial averaging to obtain more reliable weight allocation results.

[0080] Example 11: This example demonstrates the superiority of the present invention in identifying the special defect of "micro-pitting" on the surface of corrugated rolls. Pitting usually manifests as pits with a very small diameter and a certain depth on the surface, which appear as isolated dark spots in traditional two-dimensional images and are easily misjudged as stains.

[0081] Through the three-dimensional reconstruction in step S2, pitting is manifested as an inwardly pointing circularly symmetrical distribution in the normal vector field, and exhibits a local funnel-shaped feature in the height field.

[0082] In the topology construction of step S3, the nodes in the pitting region form high-gradient edge connections with their surrounding nodes and have high local curvature.

[0083] The graph neural network in step S4 can accurately distinguish pitting corrosion from ordinary surface stains by recognizing this specific "centrally symmetric anomalous topological pattern". In comparative tests, the present invention improves the recognition rate of pitting corrosion compared to traditional visual detection algorithms, and the quantized depth error is less than 2 micrometers.

[0084] Example 12: This example describes the architecture of the present invention in supporting multi-sensor fusion extension. In addition to the topographic information acquired by photometric stereo vision, the system also integrates the signal components of the acoustic emission sensor into the node features of the topology graph.

[0085] During the rotation of the corrugated roll, if stress cracks exist on the surface, ultrasonic waves of a specific frequency will be emitted. In this embodiment, this acoustic energy distribution is used as an additional node feature dimension and input into the graph neural network in step S41.

[0086] This heterogeneous fusion of physical morphology and acoustic characteristics enables the system to detect latent cracks inside corrugated rolls at an early stage. Even if some cracks have not yet extended to the surface, the subtle local vibrations they cause will produce abnormal fluctuations in the feature space of the topological map, realizing comprehensive three-dimensional monitoring of the health status of the corrugated rolls.

[0087] Example 13: This example illustrates the ultimate optimization of algorithm execution efficiency in this invention. To achieve the second-level response requirement of the corrugated paper production line, the solution logic for step S2 is embedded in a field-programmable gate array (FPGA) chip.

[0088] By implementing the least squares method and parallel pipelined processing of the parallel light intensity equations at the hardware level, the 3D reconstruction module can complete tens of millions of floating-point operations in a very short time.

[0089] The graph neural network in step S4 employs weight quantization technology, compressing the original 32-bit floating-point weights to 8-bit integers. By performing low-power inference on large-scale application-specific integrated circuits, the system reduces the power consumption per unit area of ​​corrugated roll detection while ensuring that the recognition accuracy remains essentially unchanged. This not only improves speed but also meets the energy-saving requirements of green industry.

[0090] Example 14: This example focuses on detailing the post-processing logic of the defect identification results, namely the automatic clustering and geometric feature extraction of the defect region.

[0091] After obtaining the classification probability of each node in step S43, the system uses a density-based spatial clustering algorithm to group nodes belonging to the same defect type. For each group of nodes, the system automatically fits its three-dimensional bounding box and calculates the spatial volume occupied by the defect.

[0092] For scratch-related defects, the system calculates the tortuosity parameter of the scratch by finding the shortest path in the topology graph. For wear-related defects, the system calculates the statistical variance of the height field across the entire sampling surface. These refined metrics are compiled into a standardized quality report.

[0093] The system also supports linkage with the marking machine in the workshop, spraying the corresponding QR code on the edge of the non-working area of ​​the corrugated roll. By scanning the QR code, the complete three-dimensional recognition view of that section of the corrugated roll can be retrieved on a mobile terminal, which facilitates subsequent manual review and maintenance decisions.

[0094] Example 15: This example demonstrates the universality of the invention when dealing with complex geometries (such as asymmetrical corrugated tooth profiles). Because corrugated rolls from different brands have varying tooth profile designs, traditional template-matching algorithms often fail.

[0095] This invention achieves adaptation to arbitrary tooth profiles by using a general reflection model in step S21 and employing relative geometric difference features in step S32. Graph neural networks, when processing topological structures, focus on the evolution of local geometric relationships rather than the absolute consistency of the global shape.

[0096] In a comprehensive test covering five different tooth profile corrugated rolls, the recognition accuracy of this invention fluctuated by less than 1.5%. This means that the system can seamlessly switch online on various models of corrugating machines without any hardware adjustments, demonstrating its engineering versatility and commercial promotion value.

[0097] Example 16: This example details the system's compensation mechanism in handling extreme overexposure or shadow occlusion. At the high curvature tips of the corrugated rollers, extremely strong specular highlights are often generated, leading to image pixel saturation (i.e., overexposure).

[0098] In step S22 of this embodiment, when a pixel is detected to be overexposed under the current light source, the system automatically excludes the light intensity data and instead uses the unsaturated data under the remaining light source directions for the solution. Because this invention uses light sources at eight angles, this redundancy design ensures that even in the event of local failure, an accurate normal vector solution can still be obtained.

[0099] Similarly, in the occluded areas deep within the tooth root, if a light source cannot reach them, the system will initiate interpolation filling logic based on topological correlation. The 3D reconstruction module will refer to the normal vector orientation of the surrounding unoccluded areas and combine it with the inherent geometric constraints of the corrugated roller to perform a reasonable physical deduction of the depth information of the occluded areas, ensuring the integrity of the reconstructed morphology.

[0100] This robust design enables the invention to provide stable and reliable detection output even in the face of various complex and extreme optical environments, truly solving the practicality problem in industrial settings.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for identifying surface defects on corrugated rolls, characterized in that, The method for identifying surface defects on corrugated rolls includes the following steps: S1. Obtain multi-frame two-dimensional texture images of the corrugated roller surface under illumination from multiple predetermined angle light sources through a preset lighting sequence; S2. Based on the reflected light intensity distribution information in the multi-frame two-dimensional texture images, surface normal vector inversion is performed, and the three-dimensional micro-morphological point cloud of the corrugated roll surface is reconstructed accordingly. By constructing a surface reflection model based on a predetermined reflection law, a functional mapping relationship is established between the reflected light intensity of each pixel and the surface normal vector, the light source direction vector, and the surface reflectivity. The least squares method is used to solve the light intensity equations of each pixel under multiple predetermined angle light sources, separating the surface reflectivity information and the surface normal vector information. The surface normal vector information is then spatially integrated to obtain the relative height field of the corrugated roll surface. Combined with the system calibration parameters, the three-dimensional micro-topography point cloud is generated. S3. The three-dimensional micro-topography point cloud is transformed into a topological graph structure with spatial connectivity. By downsampling the three-dimensional micro-topography point cloud, feature points reflecting surface details are retained. A set of nearest neighbor nodes for each feature point is constructed according to a preset search radius. The feature points are used as graph nodes, and the geometric neighborhood relationships determined by the set of nearest neighbor nodes are used as edges of the graph. Initial weights are assigned to the edges in the topological graph structure based on the geometric feature differences between the nearest neighbor nodes. S4. Using a graph neural network, feature extraction and message passing are performed on the topology graph structure to identify and extract abnormal topology paths representing defects, and the defect identification results are output. The node features in the topological graph structure are aggregated using graph convolutional layers. By superimposing multiple graph convolutions, topological features with long-distance dependencies are captured. The updated node features are then input into a classifier to determine the probability that the node belongs to a defect type. The defect identification result is determined based on the probability distribution.

2. The method for identifying surface defects of a corrugated roll according to claim 1, characterized in that, The predetermined reflection law mentioned in step S2 is the improved Lambert reflection law, which introduces a specular reflection correction term for highly reflective metal surfaces based on the diffuse reflection model. The specular reflection correction term controls the weight of the specular component through a preset attenuation coefficient, which is configured according to the quenching or chrome plating state of the corrugated roll surface. The surface reflection model defines the observed light intensity of each pixel as the product of the surface reflectivity, the surface normal vector, and the inner product of the light source direction vector, plus a weighted sum of the product of the specular reflection correction term and the compensation term constrained by the high brightness direction. By adjusting the weight coefficient of the specular reflection correction term, the bright spot interference caused by the high reflectivity of the metal material on the corrugated roller surface is offset.

3. The method for identifying surface defects of a corrugated roll according to claim 2, characterized in that, The process of solving for the surface normal vector information using the least squares method in step S2 includes: The acquired multi-frame two-dimensional texture images are subjected to brightness normalization processing to eliminate systematic errors caused by differences in luminous efficiency of different light sources or fluctuations in driving voltage. For each pixel location, construct a system of linear equations by combining the brightness-normalized light intensity values ​​from multiple frames with the corresponding known light source direction vector. The linear equations are optimized by using singular value decomposition to calculate the unit normal vector component of the pixel position in three-dimensional space under the constraint of minimizing the reconstruction error. The unit normal vector component includes spatial pointing data in the horizontal, vertical, and depth directions.

4. The method for identifying surface defects of a corrugated roll according to claim 3, characterized in that, The process of performing spatial integration on the surface normal vector information in step S2 includes: A variational normal integral algorithm is used to transform discontinuous surface normal vector information into a continuous height map by minimizing the difference function between the surface gradient and the measured normal vector. During the iterative process of the normal integral algorithm, a gradient operator is introduced to smooth the tooth edge of the corrugated roll. The gradient operator is configured to adjust the integral weight at the curvature abrupt changes in the tooth tip and tooth root regions of the corrugated roll, suppressing the depth distortion caused by the divergence of numerical integrals, and ensuring the spatial continuity of the relative height field under the complex tooth structure of the corrugated roll.

5. The method for identifying surface defects of a corrugated roll according to claim 4, characterized in that, The geometric feature differences mentioned in step S3 include height gradient deviation, normal angle, and local curvature variation. During the process of assigning initial weights to the edges, the angle between the normal vectors of adjacent feature points in the nearest neighbor node set is calculated, and the angle between the normal vectors is compared with a preset mutation threshold. In regions where the included angle of the normal vectors exceeds the mutation threshold, the edges are assigned initial weights to identify potential crack or scratch regions. Simultaneously, the average curvature of the feature points in the local region is calculated, and the fluctuation of the average curvature is mapped to the connection strength of the topological graph structure, thereby initially stripping the regular tooth-shaped texture features from the topological graph structure.

6. The method for identifying surface defects of a corrugated roll according to claim 5, characterized in that, The node feature aggregation process described in step S4 employs an attention mechanism; The attention mechanism is configured to automatically learn the association weights between adjacent nodes, assigning larger attention weights to adjacent nodes that conform to the defect topology distribution pattern, and assigning smaller attention weights to adjacent nodes that conform to the inherent periodic tooth shape pattern of the corrugated roll. The geometric evolution features of the topological graph structure are extracted in multiple subspaces through a multi-head attention mechanism. During the feature transfer process of the graph convolutional layer, the defect signal with random orientation features is enhanced, while the background texture signal with fixed frequency features is suppressed.

7. The method for identifying surface defects of a corrugated roll according to claim 6, characterized in that, The process of capturing topological features with long-range dependencies in step S4 includes: Residual connection structures and skip connection structures are introduced into the graph neural network to fuse the deep semantic features of the graph neural network with the original geometric features of the three-dimensional micro-topography point cloud; Global pooling is used to extract global statistical features of the topology graph structure, and the global statistical features are concatenated with local node features. The global statistical features are used to characterize the large-area wear state of the corrugated roll surface. By characterizing the physical path of crack extension along the root of the corrugated roll tooth using the long-distance dependence property, extended damage that does not have significant features in a single local view can be identified.

8. The method for identifying surface defects of a corrugated roll according to claim 7, characterized in that, After outputting the defect identification results, the method also includes a step of quantitatively evaluating the identified defects; The quantitative assessment includes calculating the maximum depth, average width, and three-dimensional extension length of the defect; for defects identified as cracks, the physical growth path of the crack is determined by searching the longest trajectory of continuous abnormal weight paths in the topological graph structure. For defects identified as wear-type defects, the degree of geometric accuracy attenuation of the surface is assessed by calculating the statistical variance of the height field within the sampling area. The defect identification results are correlated with the total amount of paper fed by the corrugated roll, and the life decay coefficient of the corrugated roll is output through the full-connection layer to predict the remaining service time of the corrugated roll.

9. A corrugated roll surface defect identification system for implementing the method as described in any one of claims 1 to 8, characterized in that, include: The image acquisition module is configured to acquire multiple frames of two-dimensional texture images of the corrugated roll surface using a high-speed industrial camera in conjunction with multiple light sources at predetermined angles. The multiple predetermined angle light sources are arranged in a ring around the high-speed industrial camera and are triggered and associated with the exposure shutter of the high-speed industrial camera through a synchronization controller; The 3D reconstruction module, whose input is connected to the image acquisition module, is used to perform surface normal vector inversion and height field integration based on photometric stereo vision, and output a 3D micro-topography point cloud. The three-dimensional reconstruction module has a built-in normal vector correction operator, which is used to eliminate and compensate for stray light intensity generated by secondary reflection based on the tooth side reflection path model of the corrugated roller. The topology construction module, whose input is connected to the three-dimensional reconstruction module, is used to convert the three-dimensional micro-topography point cloud into a topology graph structure after performing downsampling processing, and to calculate the initial weight of the edges based on the geometric feature differences between nodes. The defect identification module, whose input is connected to the topology construction module, includes a pre-trained graph neural network for receiving the topology graph structure and outputting the classification probability and spatial localization result of the defect through graph convolution operations and attention mechanisms.

Images