Surface defect detection system based on microcrystalline glass panel finished product

Abstract

The invention discloses a surface defect detection system based on a microcrystalline glass panel finished product, and relates to the technical field of surface defect detection. The defect full-dimensional dynamic state map construction module is used for constructing and updating a defect full-dimensional dynamic state map containing defect three-dimensional coordinate information, multi-modal feature confidence information, virtual-entity parameter deviation information and feature definition-registration precision linkage threshold information in real time; the structured light three-dimensional contour acquisition module is used for acquiring three-dimensional contour data of the surface of the microcrystalline glass panel; the multi-modal feature fusion module is used for collecting two-dimensional image data of multi-light-source imaging and extracting multi-modal feature information; the digital twin virtual debugging module is used for constructing a virtual detection environment based on the defect full-dimension dynamic state graph; the problem of feature confusion caused by light source combinational logic solidification in an existing multi-light-source switching system is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of surface defect detection technology, specifically a surface defect detection system based on finished microcrystalline glass panels. Background Technology

[0002] In the industrial defect detection scenario of microcrystalline glass panels, the system uses alternating imaging from multiple light sources (such as bright field, dark field and multiple wavelengths) to identify complex defects (such as scratches, ink overflow and subsurface bubbles), aiming to achieve high-precision defect classification and location.

[0003] However, the core contradiction in this scenario lies in the fact that existing multi-source switching systems use fixed preset combination logic and rigid timing triggering modes, lacking an integrated control mechanism for adaptive adjustment of source combination and dynamic calibration of switching timing. This results in the system being unable to balance the clarity of defect features and the accuracy of image registration in a dynamic detection environment.

[0004] This contradiction manifests itself in two key technical defects: First, the light source combination logic is fixed, making it impossible to dynamically optimize the light source parameters based on real-time image features, resulting in feature confusion between different types of defects (such as linear scratches and planar ink overflow) during alternating imaging; second, the switching timing control lacks feedback calibration, causing sub-pixel-level displacement accumulation between multi-light source images, resulting in ghosting phenomena during feature fusion.

[0005] These defects directly lead to increased positioning deviations and higher misjudgment rates of complex defects, seriously affecting the reliability of the detection system and its compliance with quality inspection standards. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a surface defect detection system based on finished microcrystalline glass panels.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a surface defect detection system based on finished microcrystalline glass panels, comprising: The module includes a structured light 3D contour acquisition module, a multimodal feature fusion module, a digital twin virtual debugging module, a dual closed-loop coupling control module, and a defect full-dimensional dynamic state map construction module. The defect full-dimensional dynamic state map construction module is used to construct and update in real time a defect full-dimensional dynamic state map that includes defect three-dimensional coordinate information, multimodal feature confidence information, virtual-entity parameter deviation information, and feature clarity-registration accuracy linkage threshold information. The structured light 3D contour acquisition module is used to acquire 3D contour data of the surface of the microcrystalline glass panel; The multimodal feature fusion module is used to acquire two-dimensional image data from multi-source imaging and extract multimodal feature information; The digital twin virtual debugging module is used to construct a virtual detection environment based on the full-dimensional dynamic state map of defects, simulate the imaging effect under different combinations of light source parameters, and generate an optimized combination of light source-time parameters. The dual closed-loop coupling control module is used to generate control commands based on the optimized light source-timing parameter combination, and control the light source driving unit and the camera triggering unit respectively.

[0008] As a preferred embodiment of the present invention, the structured light three-dimensional contour acquisition module includes a line structured light generator, a three-dimensional data acquisition unit, and a depth information processing unit; The line structured light generator projects a linear laser beam onto the surface of the microcrystalline glass panel; The three-dimensional data acquisition unit synchronously captures the light stripe image formed by the laser beam on the panel surface; The depth information processing unit calculates the three-dimensional coordinate data of each point on the panel surface according to the triangulation principle, and generates a three-dimensional point cloud dataset. The three-dimensional point cloud dataset contains the planar coordinate information of each data point and the depth deviation information relative to the reference plane. The depth information processing unit maps the three-dimensional point cloud dataset into a two-dimensional depth heat map according to its spatial location, which is used to distinguish between surface defect regions and subsurface defect regions.

[0009] As a preferred embodiment of the present invention, the multimodal feature fusion module includes a multi-source image acquisition unit, a grayscale feature extraction unit, an edge feature extraction unit, and a texture feature extraction unit; The multi-light source image acquisition unit acquires image data of the microcrystalline glass panel under bright field light source, dark field light source and multi-wavelength light source conditions respectively; The gray-level feature extraction unit performs adaptive histogram equalization processing on the acquired image data and calculates the mean gray level and gray-level variance of each region. The edge feature extraction unit uses an edge detection algorithm to extract edge contour information in the image and calculates the integrity index of the edge contour. The texture feature extraction unit uses a local texture analysis algorithm to calculate the texture consistency index of each region.

[0010] As a preferred embodiment of the present invention, the digital twin virtual debugging module includes a virtual environment construction unit, a light source physical model unit, and a camera imaging model unit; The virtual environment construction unit reconstructs a three-dimensional virtual model of the microcrystalline glass panel based on the three-dimensional coordinate information in the full-dimensional dynamic state map of defects; The physical model unit for the light source imports physical characteristic parameters of different types of light sources, including the wavelength range, light intensity attenuation characteristics, and illumination angle parameters. The camera imaging model unit is configured with camera parameters that match the entity detection system. These camera parameters include image resolution, optical focal length, and signal trigger delay parameters.

[0011] As a preferred embodiment of the present invention, the dual closed-loop coupling control module includes a feature sharpness control loop and a registration accuracy control loop; The feature sharpness control loop receives the simulated feature sharpness index and the target feature sharpness threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the light source parameter correction amount, and generates light source control commands. The registration accuracy control loop receives the simulated registration accuracy index and the target registration accuracy threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the trigger timing correction amount, and generates camera timing commands.

[0012] As a preferred technical solution of the present invention, the defect full-dimensional dynamic state map construction module constructs an initial state map during the system initialization phase and continuously updates the state map during system operation. The initial state map includes a three-dimensional coordinate sub-map, a multimodal feature confidence sub-map, a virtual-entity parameter deviation sub-map, and a feature clarity-registration accuracy linkage threshold sub-map; The three-dimensional coordinate sub-graph stores the depth information of each point on the panel surface; the multimodal feature confidence sub-graph stores the reliability score of the defect features in each region; The virtual-entity parameter deviation subgraph is initialized to zero and is used to record the deviation between the digital twin simulation parameters and the entity execution parameters; The feature sharpness-registration accuracy linkage threshold subgraph stores preset target feature sharpness thresholds and target registration accuracy thresholds.

[0013] As a preferred technical solution of the present invention, the surface defect detection system further includes an iterative closed-loop optimization mechanism. The iterative closed-loop optimization mechanism is configured to collect adjusted data through the structured light three-dimensional contour acquisition module and the multimodal feature fusion module after each imaging control execution, update the full-dimensional dynamic state map of the defect, and regenerate optimized control commands based on the updated map.

[0014] As a preferred embodiment of the present invention, the surface defect detection system further includes a panel positioning sensor and a synchronization control unit. The panel positioning sensor is configured to detect the edge position of the microcrystalline glass panel and output the initial position coordinate information of the panel. The synchronization control unit is configured to generate a synchronization start command to ensure the synchronization of three-dimensional data acquisition and two-dimensional image acquisition.

[0015] As a preferred embodiment of the present invention, the surface defect detection system further includes a knowledge base management module; the knowledge base management module is configured to store process parameters, defect feature databases and historical records of optimization parameters for different types of microcrystalline glass panels, and to establish a mapping relationship between panel features and optimal detection parameters.

[0016] The beneficial effects of this invention are: 1. In this invention, the depth information provided by the structured light 3D contour acquisition module and the 2D feature information extracted by the multimodal feature fusion module are organically combined to realize the multidimensional characterization of the physical properties of defects and the quality of image features in the full-dimensional dynamic state map of defects. The digital twin virtual debugging module constructs a virtual detection environment based on this map. By simulating the imaging effect under different combinations of light source parameters, it can accurately select the optimal light source-time sequence parameter combination that is suitable for specific defect features. This effectively solves the feature confusion problem caused by the solidification of light source combination logic in the existing multi-light source switching system and significantly improves the recognition accuracy of composite defects.

[0017] 2. In this invention, through the synergistic effect of the feature sharpness control loop and the registration accuracy control loop in the dual closed-loop coupled control module, the system can generate and dynamically adjust the light source control command and camera timing command in real time according to the target threshold requirements in the full-dimensional dynamic state map of the defect. This closed-loop control mechanism based on map information effectively overcomes the problem of sub-pixel-level displacement accumulation caused by the lack of feedback calibration in the switching timing control of traditional systems, ensuring accurate registration in the process of multi-source image acquisition, thereby significantly reducing the positioning deviation of micron-level defects. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a schematic diagram of the overall structure of the surface defect detection system based on microcrystalline glass panel products of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, a surface defect detection system based on a finished microcrystalline glass panel includes: The module includes a structured light 3D contour acquisition module, a multimodal feature fusion module, a digital twin virtual debugging module, a dual closed-loop coupling control module, and a defect full-dimensional dynamic state map construction module. The defect full-dimensional dynamic state map construction module is used to construct and update in real time a defect full-dimensional dynamic state map that includes defect three-dimensional coordinate information, multimodal feature confidence information, virtual-entity parameter deviation information, and feature clarity-registration accuracy linkage threshold information. The structured light 3D contour acquisition module is used to acquire 3D contour data of the surface of the microcrystalline glass panel.

[0022] The multimodal feature fusion module is used to acquire two-dimensional image data from multi-source imaging and extract multimodal feature information; The digital twin virtual debugging module is used to construct a virtual detection environment based on the full-dimensional dynamic state map of defects, simulate the imaging effect under different combinations of light source parameters, and generate an optimized combination of light source-time parameters. The dual closed-loop coupling control module is used to generate control commands based on the optimized light source-timing parameter combination, and control the light source driving unit and the camera triggering unit respectively.

[0023] In the system's workflow, the structured light 3D contour acquisition module and the multimodal feature fusion module first synchronously acquire the 3D contour data of the panel and the 2D image data of multiple light sources to construct an initial full-dimensional dynamic state map of the defect. Based on this map, the digital twin virtual debugging module tests different combinations of light source parameters in a virtual environment and selects the optimal combination of light source-time parameters. The dual closed-loop coupling control module generates control commands based on this parameter combination to drive the light source driving unit and the camera triggering unit to perform precise optical detection. During the detection process, the system continuously acquires feedback data, dynamically updates the full-dimensional dynamic state map of the defect, and re-optimizes the control parameters based on the updated map to form a closed-loop adaptive control. This architecture design breaks the traditional serial processing mode of "light source first, registration later" and realizes deep collaboration between light source control and image processing, fundamentally solving the two major technical problems of feature confusion and displacement accumulation.

[0024] Furthermore, the structured light 3D contour acquisition module includes a line structured light generator, a 3D data acquisition unit, and a depth information processing unit; The line structured light generator projects a linear laser beam onto the surface of the microcrystalline glass panel; The line structured light generator uses a high-power semiconductor laser as the light source and shapes a circular light spot into a line beam with a width of less than 0.1 mm through a cylindrical lens, which is then projected onto the surface of the microcrystalline glass panel.

[0025] The three-dimensional data acquisition unit synchronously captures the light stripe image formed by the laser beam on the panel surface; The three-dimensional data acquisition unit uses a high-precision industrial camera. During actual operation, the three-dimensional data acquisition unit maintains strict synchronization with the line structured light generator: when the line structured light generator projects a linear laser beam onto the surface of the microcrystalline glass panel, the three-dimensional data acquisition unit immediately captures the light stripe image formed by the laser beam on the panel surface. Due to defects on the surface of the microcrystalline glass panel, the light stripe will undergo local deformation. This deformation directly reflects the three-dimensional morphological features of the panel surface. In order to improve the accuracy of data acquisition, the three-dimensional data acquisition unit acquires multiple frames of images at each detection position and performs average processing, which can effectively suppress the influence of random noise.

[0026] The depth information processing unit calculates the three-dimensional coordinate data of each point on the panel surface according to the triangulation principle, and generates a three-dimensional point cloud dataset. The three-dimensional point cloud dataset contains the planar coordinate information of each data point and the depth deviation information relative to the reference plane. The depth information processing unit calculates the three-dimensional coordinate data of each point on the panel surface based on the triangulation principle (which is a mature technology in the field of three-dimensional optical measurement and will not be elaborated here).

[0027] The depth information processing unit maps the three-dimensional point cloud dataset into a two-dimensional depth heat map according to the spatial location, which is used to distinguish between surface defect regions and subsurface defect regions. Two-dimensional depth thermal mapping plays a crucial role in defect identification. Based on the manufacturing process and defect characteristics of the microcrystalline glass panel, surface defect areas (such as scratches and ink overflow) are usually located on the surface of the panel with small depth deviations, generally within the range of -1μm to 1μm. Subsurface defect areas (such as bubbles and inclusions) are located inside the panel, and their surface manifestations are local bulges or depressions with larger depth deviations, generally within the range of 1μm to 20μm. By setting a depth threshold range, the depth information processing unit can automatically distinguish different types of defect areas, providing a key basis for subsequent optimization of light source parameters.

[0028] Through this design, the structured light 3D contour acquisition module not only provides depth information that traditional 2D detection systems cannot obtain, but more importantly, it establishes the correlation between surface morphology and optical properties, laying the physical foundation for an integrated control mechanism that enables adaptive adjustment of light source combination and dynamic calibration of switching timing.

[0029] Furthermore, the multimodal feature fusion module includes a multi-source image acquisition unit, a grayscale feature extraction unit, an edge feature extraction unit, and a texture feature extraction unit; The multi-light source image acquisition unit acquires image data of the microcrystalline glass panel under bright field light source, dark field light source and multi-wavelength light source conditions respectively; The gray-level feature extraction unit performs adaptive histogram equalization processing on the acquired image data and calculates the mean gray level and gray-level variance of each region. Traditional histogram equalization is prone to over-enhancing noise when processing high dynamic range images, while adaptive histogram equalization effectively solves this problem by dividing the image into multiple local regions and performing histogram equalization independently in each region.

[0030] For example, in a specific implementation, the image is divided into overlapping regions of 64×64 pixels. The histogram equalization result of each region is smoothly transitioned through bilinear interpolation. In the processed image, the mean gray value and gray variance of each pixel region are accurately calculated.

[0031] The grayscale mean reflects the overall brightness level of the area, while the grayscale variance characterizes the contrast characteristics of the area. In the inspection of microcrystalline glass panels, ink overflow defects usually manifest as a significant change in the local grayscale mean, while scratch defects are more often manifested as an increase in the local grayscale variance.

[0032] The edge feature extraction unit uses an edge detection algorithm (Canny algorithm is a classic edge detection algorithm) to extract edge contour information in the image and calculates the integrity index of the edge contour; The texture feature extraction unit uses a local texture analysis algorithm to calculate the texture consistency index of each region. Optionally, the texture feature extraction unit uses the Local Binary Pattern (LBP) algorithm to calculate the texture consistency index of each region. The LBP algorithm extracts texture features by comparing the gray value relationship within the pixel neighborhood and has good robustness to illumination changes.

[0033] The multimodal feature fusion module performs weighted fusion of grayscale features, edge features, and texture features to generate a multimodal feature confidence map, which is used to characterize the reliability of defect features in different regions.

[0034] Specifically, this invention proposes a dynamic weight adjustment mechanism based on defect type, which achieves adaptive feature fusion through a multimodal feature confidence calculation algorithm. The mathematical expression of this algorithm is as follows: ; Where Mc(x,y) represents the confidence level of the defect feature in the (x,y) pixel region, with a value range of [0,1]. The larger the value, the more obvious the defect feature in the region. ω1 is the weight coefficient of grayscale contrast; ω2 is the weight coefficient of edge integrity; ω3 is the weight coefficient of texture consistency; C(x,y) is the grayscale contrast, which is the normalized result of the grayscale difference between the (x,y) region and the background region, with a value range of [0,1]; E(x,y) is the edge integrity, which is the ratio of the continuous length of the defect edge in the (x,y) region to the expected defect edge length, with a value range of [0,1]; T(x,y) is the texture consistency, which is the matching degree between the texture of the (x,y) region and the typical texture of the corresponding defect type, with a value range of [0,1].

[0035] For example, for linear defects (such as scratches): the core feature is edge integrity, and ω2=0.5, ω1=0.3, ω3=0.2 are set to highlight the weight of edge features; For planar defects (such as ink overflow): the core feature is texture consistency, and ω3=0.5, ω1=0.4, and ω2=0.1 are set to highlight the weight of texture features; When the weight of the core feature is ≥0.5, it can effectively distinguish the feature differences of different defects. The weight of the non-core feature is ≥0.1, which can help correct the identification error and avoid misjudgment caused by the failure of a single feature.

[0036] Furthermore, the digital twin virtual debugging module includes a virtual environment construction unit, a light source physical model unit, and a camera imaging model unit; The virtual environment construction unit reconstructs a three-dimensional virtual model of the microcrystalline glass panel based on the three-dimensional coordinate information in the full-dimensional dynamic state map of defects; The virtual environment construction unit is a fundamental component of the digital twin virtual debugging module. It is responsible for converting the three-dimensional coordinate information in the full-dimensional dynamic state map of the defect into a visualized three-dimensional virtual model. During the initialization phase, this unit receives the initial state map from the full-dimensional dynamic state map construction module of the defect and extracts the three-dimensional coordinate sub-map. The three-dimensional coordinate sub-map stores the depth information of each point on the surface of the microcrystalline glass panel. These data originate from the high-precision three-dimensional point cloud data collected by the structured light three-dimensional contour acquisition module. The virtual environment construction unit adopts an advanced surface reconstruction algorithm to convert the discrete point cloud data into a continuous triangular mesh surface model. The mesh density can be dynamically adjusted according to the detection accuracy requirements to ensure that the virtual model can accurately reflect the microscopic morphological features of the panel surface. The virtual environment construction unit also includes a material property configuration module, which sets corresponding optical parameters in the virtual model based on the physical properties of the microcrystalline glass panel, including surface reflectivity, refractive index, transmittance, and scattering coefficient.

[0037] The light source physical model unit imports physical characteristic parameters of different types of light sources, including light source wavelength range, light intensity attenuation characteristics and illumination angle parameters, and simulates the propagation, reflection and scattering behavior of light on the virtual panel surface based on the physics engine; The light source physical model unit is the core computing engine of the digital twin virtual debugging module, responsible for simulating the optical behavior of different types of light sources on the virtual panel surface. This unit first imports the physical characteristic parameters of bright field light sources, dark field light sources and multi-wavelength light sources. These parameters include the wavelength range of the light source, light intensity attenuation characteristics and illumination angle parameters. All parameters are calibrated based on the measured data of physical light source devices. The light source physical model unit adopts physically based rendering technology (PBR) to accurately simulate the propagation, reflection and scattering behavior of light on the surface of microcrystalline glass.

[0038] The camera imaging model unit is configured with camera parameters that match the entity detection system. These camera parameters include image resolution, optical focal length, and signal trigger delay parameters. The camera imaging model unit is responsible for simulating the imaging process of a physical camera, ensuring the consistency of the virtual image and the physical image in terms of geometric and optical characteristics. This unit first sets camera parameters that are completely matched with the physical detection system, including image resolution, optical focal length, and signal trigger delay parameters. The image resolution is set to be consistent with the sensor resolution of the physical camera, ensuring that the virtual image and the physical image correspond completely in spatial scale. The optical focal length parameter is set based on the actual specifications of the camera lens to accurately simulate the perspective effect of the imaging. The signal trigger delay parameter is a key parameter affecting the registration accuracy of multi-source image. This parameter is set based on the measured response time of the physical camera and is used to simulate the time delay from receiving the trigger signal to the start of exposure. During the imaging simulation, the camera imaging model unit considers various factors affecting image quality, including lens distortion, depth of field effect, motion blur, and sensor noise. For lens distortion, the system uses a pinhole camera model combined with radial distortion coefficients for correction. For depth of field effect, the system simulates changes in image sharpness at different object distances, which is particularly important in 3D panel inspection. For motion blur, the system considers the relationship between panel transmission speed and exposure time to accurately simulate the degree of image blur during motion. For sensor noise, the system adds random noise that conforms to actual distribution, making the virtual image closer to the characteristics of physical imaging.

[0039] The digital twin virtual debugging module traverses different combinations of light source parameters in a virtual environment, simulates the imaging effect under various parameter combinations, evaluates the feature sharpness index and registration accuracy index of the simulated image, and selects the optimal light source-time parameter combination that meets the preset threshold requirements.

[0040] The parameter traversal range includes light source wavelength (covering the visible spectrum range), light intensity (from the lowest detectable intensity to the maximum safe intensity), and trigger timing (the theoretical range calculated based on the panel transmission speed). For each set of parameters, the system executes a complete optical simulation process: first, the virtual environment construction unit loads the current panel model; then, the light source physical model unit applies the current light source parameters to simulate light; and finally, the camera imaging model unit generates a virtual image.

[0041] During the parameter evaluation phase, the system performs a dual evaluation on the generated virtual image: on the one hand, it calculates the feature sharpness index of the simulated image, which reflects the obviousness of the defect features in the image and is obtained by comparing the similarity with the confidence value in the multimodal feature confidence sub-map; on the other hand, it calculates the registration accuracy index of the simulated image, which reflects the spatial consistency between images under different light source conditions and is obtained by calculating the average displacement deviation of feature points. These two indices directly correspond to the preset threshold in the feature sharpness-registration accuracy linkage threshold sub-map in the full-dimensional dynamic state map of defects.

[0042] Specifically, the virtual-entity parameter deviation algorithm formula in this invention is as follows: ; ; Among them, P sim_new Represented as the corrected light source simulation parameters (light intensity / wavelength); P sim_old T represents the light source simulation parameters from the previous round; sim_new Represented as the corrected timing simulation parameters (trigger interval); T sim_old The parameters represent the timing simulation parameters from the previous round; k represents the correction coefficient, with a value range of (0,1); ΔP represents the virtual-to-physical deviation of the light source parameters, i.e., P sim_old -P real P real These are the actual light source parameters; ΔT represents the virtual-to-physical deviation of the time series parameters, i.e., Ts. im_old -T real T real This refers to the actual trigger timing.

[0043] In this embodiment, the preferred value of the correction coefficient k is 0.8. When k=0.8, it can ensure the correction speed (the current deviation can be reduced by 80% in each iteration) and avoid overcorrection (20% of the deviation is left for subsequent fine-tuning to prevent system oscillation). If k≤0.5, the correction speed is too slow and cannot meet the real-time requirements of industrial detection. If k≥0.9, the "overshoot" phenomenon is likely to occur, causing the parameters to fluctuate around the optimal value and affecting the detection stability. In addition, the reasonable range of light source parameter deviation ΔP is: light intensity deviation ≤ 50cd, wavelength deviation ≤ 20nm; and the reasonable range of timing parameter deviation ΔT is ≤ 5ms.

[0044] The parameter selection process adopts a multi-objective optimization strategy, prioritizing parameter combinations that simultaneously meet the feature sharpness threshold and registration accuracy threshold. When multiple combinations meet the conditions, the system uses a weighted scoring mechanism to comprehensively consider the degree of compliance of the two indicators and select the combination with the highest comprehensive score as the optimal light source-time series parameter combination.

[0045] This screening mechanism ensures that the generated parameter combinations not only perform well on individual indicators, but also achieve an optimal balance in overall detection performance.

[0046] In actual operation, the digital twin virtual debugging module and the defect full-dimensional dynamic state map construction module form a closed-loop interaction. When the system receives the updated state map, especially the virtual-entity parameter deviation sub-map, the digital twin virtual debugging module will automatically correct the virtual model parameters.

[0047] Furthermore, the dual closed-loop coupling control module includes a feature sharpness control loop and a registration accuracy control loop; The feature sharpness control loop receives the simulated feature sharpness index and the target feature sharpness threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the light source parameter correction amount, and generates light source control commands. The core function of the feature sharpness control loop is to ensure that images acquired under different light source conditions have sufficient feature sharpness so that defect features can be accurately identified and classified. The input data of this loop includes two key information sources: one is the simulated feature sharpness index output by the digital twin virtual debugging module, which reflects the expected image feature quality under a specific combination of light source parameters in the virtual environment; the other is the target feature sharpness threshold from the full-dimensional dynamic state map of the defect, which represents the minimum image quality requirement of the current detection task.

[0048] During the control calculation process, the feature sharpness control loop first compares the difference between the simulated feature sharpness index and the target feature sharpness threshold. When the simulated feature sharpness index is lower than the target feature sharpness threshold, it indicates that the current combination of light source parameters cannot provide sufficient image quality and parameter correction is required. The correction amount is calculated using a proportional control strategy, that is, the amount of light source parameter correction is proportional to the difference in feature sharpness. The proportional coefficient is dynamically adjusted according to the historical control effect. This control strategy ensures the smoothness and stability of parameter adjustment and avoids unstable imaging quality caused by drastic fluctuations in light source parameters.

[0049] Specifically, the target feature sharpness threshold is calculated as follows: ; Among them, St new St represents the updated target feature sharpness threshold, with a value range of [0,1]. old S represents the target feature sharpness threshold of the previous round, α is the adjustment coefficient, and its value ranges from (0,1). target This represents the final target feature sharpness, with a preset value of 0.8, S. real Represented as actual feature clarity, and is the mean of the multimodal feature confidence subplot; Preferably, α=0.3. The adjustment of feature sharpness needs to take into account both accuracy and stability. An adjustment coefficient of 0.3 can make the threshold update by 30% in each iteration, avoiding large adjustments to the light source parameters due to sudden changes in the threshold, and ensuring stable image quality.

[0050] The registration accuracy control loop receives the simulated registration accuracy index and the target registration accuracy threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the trigger timing correction amount, and generates camera timing commands. The core function of the registration accuracy control loop is to ensure spatial consistency between multi-source images and eliminate sub-pixel-level displacement accumulation caused by factors such as mechanical vibration and transmission speed fluctuations. The input data of this loop also includes two key information sources: one is the simulated registration accuracy index output by the digital twin virtual debugging module, which reflects the expected image registration quality under specific time parameters in the virtual environment; the other is the target registration accuracy threshold from the defect full-dimensional dynamic state map, which represents the strict requirements of the current detection task for spatial consistency.

[0051] During the control calculation process, the registration accuracy control loop first evaluates the difference between the simulated registration accuracy index and the target registration accuracy threshold. When the simulated registration accuracy index is worse than the target registration accuracy threshold, it indicates that the current timing parameters cannot meet the registration requirements and timing correction is required. The calculation of the correction amount also adopts a proportional control strategy, but takes into account the dynamic characteristics and time delay characteristics of the system and introduces a feedforward compensation mechanism. That is, it predicts the future displacement deviation based on the changing trend of the panel transmission speed and adjusts the trigger timing in advance. This forward-looking control strategy significantly improves the system's response speed and control accuracy.

[0052] Specifically, the target registration accuracy threshold is calculated as follows: ; Among them, Et new Et represents the updated target registration accuracy threshold, in μm. oldE represents the target registration accuracy threshold of the previous round, β is the adjustment coefficient, and its value ranges from (0,1). target E represents the final target registration accuracy, with a preset value of 0.1 μm. real This represents the actual registration accuracy, which is the average value of the feature point displacement deviation between images from different light sources. Preferably, β=0.2: Registration accuracy has a greater impact on the detection results and requires a smoother adjustment. An adjustment coefficient of 0.2 can reduce the fluctuation of timing parameters and avoid displacement accumulation.

[0053] The dual-closed-loop coupled control module transmits light source control commands to the light source driving unit and camera timing commands to the camera triggering unit via an industrial real-time communication network. The light source driving unit adjusts the wavelength and light intensity parameters of the light source according to the light source control commands. The camera triggering unit adjusts the triggering time interval for image acquisition according to the camera timing commands.

[0054] Image quality and spatial consistency are often considered independent issues, requiring different optimization strategies. However, in practical applications of microcrystalline glass panel inspection, these two issues are inherently coupled: improving feature sharpness may require increasing exposure time or adjusting the light source angle, but this exacerbates displacement accumulation; improving registration accuracy may require shortening exposure time or reducing the light source angle, but this reduces feature sharpness. This coupling means that optimizing one objective often leads to the deterioration of the other, creating a vicious cycle.

[0055] This invention establishes a dynamic balance mechanism between two control loops by using a feature sharpness-registration accuracy linkage threshold subgraph in the full-dimensional dynamic state map of defects. This linkage threshold subgraph stores the weight requirements of the two indicators for the current detection task. When the system detects a composite defect, it automatically increases the weight of feature sharpness; when it detects a micron-level defect, it automatically increases the weight of registration accuracy. This dynamic weight allocation mechanism enables the dual closed-loop control to automatically adjust the optimization strategy under different detection scenarios, achieving global optimization rather than local optimization.

[0056] The multimodal feature confidence calculation algorithm provides defect feature reliability data Mc(x,y) for subsequent algorithms. The digital twin virtual debugging module calculates the simulated feature sharpness based on Mc(x,y), and the dual closed-loop coupled control module calculates the actual feature sharpness S based on the mean of Mc(x,y). real ; The virtual-entity parameter deviation algorithm formula is based on the feature data provided by the multimodal feature confidence calculation algorithm and the threshold data provided by the algorithm formula of target feature clarity threshold and target registration accuracy threshold. It corrects the simulation parameters of the digital twin, making the parameters closer to the real scene, and provides accurate initial parameters for dual closed-loop control. The algorithm formulas for the target feature sharpness threshold and the target registration accuracy threshold are based on the actual feature sharpness S calculated by the multimodal feature confidence algorithm. real The effect of parameter execution after correction of the virtual-entity parameter deviation algorithm formula (actual registration accuracy S) real The threshold is dynamically adjusted to provide suitable target requirements for parameter simulation and dual-loop control execution of digital twins.

[0057] The three elements work together to ensure the optimization of the entire process of the system, from feature recognition to parameter adjustment and threshold adaptation, thereby achieving a synergistic improvement in detection accuracy and efficiency.

[0058] Furthermore, the defect full-dimensional dynamic state map construction module constructs an initial state map during the system initialization phase and continuously updates the state map during system operation. The initial state map includes a three-dimensional coordinate sub-map, a multimodal feature confidence sub-map, a virtual-entity parameter deviation sub-map, and a feature clarity-registration accuracy linkage threshold sub-map.

[0059] The three-dimensional coordinate sub-graph stores the depth information of each point on the panel surface; The three-dimensional coordinate sub-graph is the physical basis of the initial state map, storing the depth information of each point on the surface of the microcrystalline glass panel. This depth data originates from the three-dimensional point cloud data acquired by the structured light three-dimensional contour acquisition module. The discrete point cloud data is converted into a two-dimensional depth distribution consistent with the camera image resolution through a spatial mapping algorithm. In the inspection of microcrystalline glass panels, depth information is of decisive significance: surface defect areas (such as scratches and ink overflow) are usually located on the surface of the panel with small depth deviations; while subsurface defect areas (such as bubbles and inclusions) are located inside the panel, and their surface manifestations are local bulges or depressions with larger depth deviations. Through accurate depth information, the system can distinguish different types of defects from a physical perspective, providing a key basis for subsequent optimization of light source parameters.

[0060] The multimodal feature confidence subgraph stores the reliability score of the defect features in each region; The multimodal feature confidence subgraph stores the reliability scores of defect features in each region. These scores are derived from the multidimensional feature information extracted by the multimodal feature fusion module. The core value of this subgraph lies in quantifying the clarity of defect features in different regions, enabling the system to identify high-confidence regions and low-confidence regions.

[0061] In complex defect detection scenarios, certain areas may simultaneously contain multiple defect types, which traditional methods often struggle to distinguish. However, by using multimodal feature confidence scoring, the system can identify the dominant feature regions of different defects, providing a decision-making basis for targeted optimization of light source parameters. The value range of this sub-image is 0-1, with values ​​closer to 1 indicating more defined defect features and values ​​closer to 0 indicating more ambiguous or uncertain features.

[0062] The virtual-entity parameter deviation subgraph is initialized to zero and is used to record the deviation between the digital twin simulation parameters and the entity execution parameters; This is the starting point of the system's self-learning mechanism. This subgraph records the deviations between the digital twin simulation parameters and the entity's execution parameters, including deviations in light source parameters and timing parameters. When the system first runs, due to a lack of historical data, the deviations between the virtual model and the entity system are unknown, so they are initialized to zero. As the system runs, this subgraph gradually accumulates deviation data, forming a deviation distribution map, providing a basis for continuous correction of the virtual model. This design reflects the system's incremental learning capability, eliminating the need for pre-calibrating a large number of parameters, but automatically optimizing through actual operating data.

[0063] The feature sharpness-registration accuracy linkage threshold sub-graph stores preset target feature sharpness thresholds and target registration accuracy thresholds; These thresholds are preset based on the type and process characteristics of the glass-ceramic panel. The unique design of this sub-graph lies in linking two traditionally independent indicators, reflecting the invention's profound understanding of system coupling.

[0064] In the inspection of microcrystalline glass panels, feature sharpness and registration accuracy often have an inverse relationship: improving feature sharpness may require a longer exposure time or a specific light source angle, but this will exacerbate displacement accumulation and reduce registration accuracy; improving registration accuracy may require a shorter exposure time or a smaller light source angle, but this will reduce feature sharpness. By setting a linkage threshold, the system can find the best balance between the two, rather than pursuing the best of a single indicator.

[0065] During system operation, the defect full-dimensional dynamic state map construction module receives real-time data from each module and dynamically updates the content of each sub-map, providing decision-making basis for the digital twin virtual debugging module and the dual closed-loop coupled control module. This enables the system to respond in real time to changes in the environment and detection requirements. The update process adopts an incremental update strategy, modifying only the data areas that have changed, rather than redrawing the entire map, which significantly improves update efficiency and system real-time performance.

[0066] Furthermore, the surface defect detection system also includes an iterative closed-loop optimization mechanism. The iterative closed-loop optimization mechanism is configured to collect adjusted data through the structured light three-dimensional contour acquisition module and the multimodal feature fusion module after each imaging control execution, update the full-dimensional dynamic state map of the defect, and regenerate optimized control commands based on the updated map.

[0067] The iterative closed-loop optimization mechanism adopts a hierarchical control architecture, which includes four functional layers: data acquisition layer, map update layer, parameter optimization layer, and instruction generation layer. Each layer is seamlessly connected through a standardized data interface.

[0068] The data acquisition layer is responsible for acquiring the adjusted 3D contour data and 2D feature data; the map update layer is responsible for integrating the acquired data into the full-dimensional dynamic state map of the defect; the parameter optimization layer corrects the virtual model and regenerates the parameters based on the updated map; and the instruction generation layer is responsible for converting the optimized parameters into actual control instructions.

[0069] This layered architecture design ensures the modularity and scalability of the iteration process, while guaranteeing the clarity of data flow and the traceability of control logic.

[0070] At the data acquisition level, the iterative closed-loop optimization mechanism works collaboratively with the structured light 3D contour acquisition module and the multimodal feature fusion module to obtain high-precision feedback data. The structured light 3D contour acquisition module performs a secondary scan of the same detection area under adjusted light source and timing parameters, acquiring updated 3D point cloud data. This secondary scan is not a simple repetition of measurement, but rather a targeted optimization of scanning parameters based on the depth benchmark established in the first scan, particularly increasing the scanning density in low-confidence areas to improve the accuracy of depth measurement. The multimodal feature fusion module, under the optimized light source-timing parameter combination, re-acquires multi-light source images and extracts updated 2D feature vectors. These feature vectors not only contain basic features such as grayscale, edges, and textures, but also include information on the correlation between features, providing a comprehensive basis for evaluating the parameter optimization effect.

[0071] At the map update level, the defect full-dimensional dynamic state map construction module performs fine-grained updates to the defect full-dimensional dynamic state map based on the collected adjusted data. The update process adopts an incremental strategy, modifying only the data areas that have changed, rather than redrawing the entire map. For the three-dimensional coordinate sub-map, the update is based on the depth data obtained from the second scan. By comparing the old and new depth values, the actual depth deviation is calculated, and the distribution of the depth heatmap is adjusted accordingly. For the multimodal feature confidence sub-map, the update is based on the feature quality extracted from the adjusted image. By calculating the actual feature clarity, the confidence score of each region is dynamically adjusted. Particularly important is the update of the virtual-entity parameter deviation sub-map, which records the differences between the digital twin simulation parameters and the entity execution parameters. These difference data are key inputs for virtual model correction.

[0072] At the parameter optimization level, the corrected virtual model regenerates an optimized combination of light source and time sequence parameters. This process is not a parameter search from scratch, but rather an intelligent guidance based on historical optimization trajectories. The system retains the parameter combinations and their effect evaluations from previous iterations. By analyzing this historical data, the system identifies the mapping relationship between parameter changes and detection effects, thereby locating the optimal solution more efficiently in the parameter space. Especially when dealing with complex defect areas, the system generates differentiated parameter combinations based on the characteristic response characteristics of different defect types, avoiding the problem that a single parameter combination cannot simultaneously optimize the detection effects of multiple defects.

[0073] At the iterative control level, the iterative closed-loop optimization mechanism sets scientific iterative termination conditions, which include two main dimensions: first, performance attainment conditions, that is, after multiple consecutive iterations, the actual feature clarity reaches the target threshold and the actual registration accuracy meets the target requirements; second, iteration count limits, that is, when the preset maximum number of iterations is reached, the iteration is forcibly terminated to avoid infinite loops. Performance attainment conditions ensure detection quality, while iteration count limits ensure system real-time performance. The combination of the two achieves the best balance between quality and efficiency.

[0074] Furthermore, the surface defect detection system also includes a panel positioning sensor and a synchronization control unit. The panel positioning sensor is configured to detect the edge position of the microcrystalline glass panel and output the initial position coordinate information of the panel. The core function of this sensor is not only to detect the position of the panel edge, but more importantly, to provide a precise spatiotemporal reference for the entire detection system. In its specific implementation, the panel positioning sensor contains multiple photoelectric sensing elements, which are arranged along the panel transmission direction to form a position detection array. When the edge of the microcrystalline glass panel passes through the detection area, the photoelectric sensing elements are triggered in sequence, and the position coordinates of the panel edge are accurately determined by time difference calculation.

[0075] The synchronization control unit is configured to generate a synchronization start command to ensure synchronization between 3D data acquisition and 2D image acquisition.

[0076] This unit employs a hybrid synchronization architecture that combines hardware triggering and software coordination, ensuring both synchronization accuracy and system flexibility. Upon receiving the trigger signal from the panel positioning sensor, the synchronization control unit first verifies the signal validity, eliminating false triggers and noise interference, and then generates a high-precision synchronization start command.

[0077] Furthermore, the system also includes a knowledge base management module; the knowledge base management module is configured to store process parameters, defect feature libraries and historical records of optimization parameters for different types of microcrystalline glass panels, and establish a mapping relationship between panel features and optimal detection parameters.

[0078] The knowledge base management module not only solves the technical problem of long debugging cycle of new panels mentioned in the background technology, but more importantly, it realizes the continuous evolution and self-optimization of the detection system through knowledge accumulation and intelligent application, providing sustainable technical guarantee for high-quality and high-efficiency detection of microcrystalline glass panels.

[0079] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A surface defect detection system based on finished microcrystalline glass panels, characterized in that, include: The module includes a structured light 3D contour acquisition module, a multimodal feature fusion module, a digital twin virtual debugging module, a dual closed-loop coupling control module, and a defect full-dimensional dynamic state map construction module. The defect full-dimensional dynamic state map construction module is used to construct and update in real time a defect full-dimensional dynamic state map that includes defect three-dimensional coordinate information, multimodal feature confidence information, virtual-entity parameter deviation information, and feature clarity-registration accuracy linkage threshold information. The structured light 3D contour acquisition module is used to acquire 3D contour data of the surface of the microcrystalline glass panel; The multimodal feature fusion module is used to acquire two-dimensional image data from multi-source imaging and extract multimodal feature information; The digital twin virtual debugging module is used to construct a virtual detection environment based on the full-dimensional dynamic state map of defects, simulate the imaging effect under different combinations of light source parameters, and generate an optimized combination of light source-time parameters. The dual closed-loop coupling control module is used to generate control commands based on the optimized light source-timing parameter combination, and control the light source driving unit and the camera triggering unit respectively.

2. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The structured light 3D contour acquisition module includes a line structured light generator, a 3D data acquisition unit, and a depth information processing unit. The line structured light generator projects a linear laser beam onto the surface of the microcrystalline glass panel; The three-dimensional data acquisition unit synchronously captures the light stripe image formed by the laser beam on the panel surface; The depth information processing unit calculates the three-dimensional coordinate data of each point on the panel surface according to the triangulation principle, and generates a three-dimensional point cloud dataset. The three-dimensional point cloud dataset contains the planar coordinate information of each data point and the depth deviation information relative to the reference plane. The depth information processing unit maps the three-dimensional point cloud dataset into a two-dimensional depth heat map according to its spatial location, which is used to distinguish between surface defect regions and subsurface defect regions.

3. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The multimodal feature fusion module includes a multi-source image acquisition unit, a grayscale feature extraction unit, an edge feature extraction unit, and a texture feature extraction unit; The multi-light source image acquisition unit acquires image data of the microcrystalline glass panel under bright field light source, dark field light source and multi-wavelength light source conditions respectively; The gray-level feature extraction unit performs adaptive histogram equalization processing on the acquired image data and calculates the mean gray level and gray-level variance of each region. The edge feature extraction unit uses an edge detection algorithm to extract edge contour information in the image and calculates the integrity index of the edge contour. The texture feature extraction unit uses a local texture analysis algorithm to calculate the texture consistency index of each region.

4. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The digital twin virtual debugging module includes a virtual environment construction unit, a light source physical model unit, and a camera imaging model unit; The virtual environment construction unit reconstructs a three-dimensional virtual model of the microcrystalline glass panel based on the three-dimensional coordinate information in the full-dimensional dynamic state map of defects; The physical model unit for the light source imports physical characteristic parameters of different types of light sources, including the wavelength range, light intensity attenuation characteristics, and illumination angle parameters. The camera imaging model unit is configured with camera parameters that match the entity detection system. These camera parameters include image resolution, optical focal length, and signal trigger delay parameters.

5. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The dual closed-loop coupling control module includes a feature sharpness control loop and a registration accuracy control loop; The feature sharpness control loop receives the simulated feature sharpness index and the target feature sharpness threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the light source parameter correction amount, and generates light source control commands. The registration accuracy control loop receives the simulated registration accuracy index and the target registration accuracy threshold in the full-dimensional dynamic state map of defects output by the digital twin virtual debugging module, calculates the trigger timing correction amount, and generates camera timing commands.

6. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The defect full-dimensional dynamic state map construction module constructs an initial state map during the system initialization phase and continuously updates the state map during system operation. The initial state map includes a three-dimensional coordinate sub-map, a multimodal feature confidence sub-map, a virtual-entity parameter deviation sub-map, and a feature clarity-registration accuracy linkage threshold sub-map; The three-dimensional coordinate sub-graph stores the depth information of each point on the panel surface; the multimodal feature confidence sub-graph stores the reliability score of the defect features in each region; The virtual-entity parameter deviation subgraph is initialized to zero and is used to record the deviation between the digital twin simulation parameters and the entity execution parameters; The feature sharpness-registration accuracy linkage threshold subgraph stores preset target feature sharpness thresholds and target registration accuracy thresholds.

7. A surface defect detection system based on a microcrystalline glass panel as described in claim 6, characterized in that, The surface defect detection system also includes an iterative closed-loop optimization mechanism. The iterative closed-loop optimization mechanism is configured to collect adjusted data through the structured light three-dimensional contour acquisition module and the multimodal feature fusion module after each imaging control execution, update the full-dimensional dynamic state map of the defect, and regenerate optimized control commands based on the updated map.

8. The surface defect detection system based on a microcrystalline glass panel as described in claim 1, characterized in that, The surface defect detection system also includes a panel positioning sensor and a synchronization control unit. The panel positioning sensor is configured to detect the edge position of the microcrystalline glass panel and output the initial position coordinate information of the panel. The synchronization control unit is configured to generate a synchronization start command to ensure the synchronization of three-dimensional data acquisition and two-dimensional image acquisition.

9. A surface defect detection system based on a microcrystalline glass panel as described in claim 4, characterized in that, The surface defect detection system also includes a knowledge base management module; the knowledge base management module is configured to store process parameters, defect feature databases and historical records of optimized parameters for different types of microcrystalline glass panels, and establish a mapping relationship between panel features and optimal detection parameters.

Images