Power copper bar production quality visual inspection system and method

Abstract

The invention provides a visual inspection system and method for the production quality of an electric power copper bar, and relates to the technical field of visual inspection, and the system comprises a multi-mode imaging module which is used for obtaining a polarization high-dynamic surface image, a laser scanning three-dimensional point cloud and a microspur infrared dual-mode image; the space registration module is used for carrying out CAD model space registration to obtain a sub-pixel aligned three-dimensional point cloud; the defect attribution detection module is used for carrying out defect attribution detection to obtain a defect process cause and effect diagram; the dead zone verification module is used for performing electrical safety dead zone verification to obtain a safety margin index; and the reverse traceability module is used for reverse traceability of production defects and positioning defect source process nodes. The technical problem that the defect detection of the power copper bar depends on a single visual imaging technology, surface defects and three-dimensional geometric information cannot be comprehensively captured, omission or misjudgment of product quality detection is caused, and the production quality of the power copper bar is affected in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, specifically to a visual inspection system and method for the production quality of power copper busbars. Background Technology

[0002] Copper busbars are important conductive components in power systems and are widely used in power transmission and distribution equipment. The quality of copper busbars directly affects the stability and safety of the entire power system, so quality inspection during their production process is crucial.

[0003] Most existing technologies use ordinary visible light imaging or simple 3D scanning systems to detect surface defects in power copper busbars. These single imaging techniques can only detect surface defects such as scratches and bubbles, and cannot fully capture the three-dimensional shape of the copper busbar or more complex geometric defects. For example, surface defects may be difficult to detect due to the reflection of the plating, or three-dimensional geometric defects such as deformation and bulges cannot be captured by conventional two-dimensional images. Therefore, detection methods that rely solely on a single imaging method are prone to omissions and misjudgments in product quality inspection, thereby affecting the production quality of power copper busbars. Summary of the Invention

[0004] This application provides a visual inspection system and method for the production quality of power copper busbars, aiming to solve the technical problem that the existing technology for defect detection of power copper busbars relies on a single visual imaging technology, which cannot fully capture surface defects and three-dimensional geometric information, leading to omissions or misjudgments in product quality inspection, thereby affecting the production quality of power copper busbars.

[0005] The first aspect disclosed in this application provides a visual inspection system for the production quality of power copper busbars. The system includes: a multimodal imaging module for driving a composite inspection station deployed at the end of the production line to simultaneously perform multimodal imaging on the power copper busbars to be inspected conveyed by a conveyor belt, obtaining a polarized high dynamic range surface image, a laser-scanned 3D point cloud, and a macro-infrared dual-mode image; a spatial registration module for constructing a spatial transformation matrix based on the center feature points of the mounting holes of the power copper busbars to be inspected using an iterative nearest-point algorithm, and performing CAD model spatial registration on the laser-scanned 3D point cloud to obtain a sub-pixel aligned 3D point cloud; a defect attribution detection module for inputting the polarized high dynamic range surface image and the sub-pixel aligned 3D point cloud into a cross-modal defect attribution network for defect attribution detection, obtaining a defect process cause-effect graph; a dead-zone verification module for performing electrical safety dead-zone verification based on the macro-infrared dual-mode image to obtain a safety margin index; and a reverse tracing module for performing reverse tracing of production defects based on the defect process cause-effect graph and the safety margin index to locate the defect source process node.

[0006] The second aspect of this application discloses a visual inspection method for the production quality of power copper busbars. This method is implemented using the aforementioned visual inspection system for power copper busbar production quality. The method includes: driving a composite inspection station deployed at the end of the production line to perform synchronous multimodal imaging on the power copper busbar to be inspected conveyed by a conveyor belt, obtaining a polarized high dynamic range surface image, a laser-scanned three-dimensional point cloud, and a macro-infrared dual-mode image; constructing a spatial transformation matrix based on the center feature points of the mounting holes of the power copper busbar to be inspected using an iterative nearest-point algorithm, and performing CAD model spatial registration on the laser-scanned three-dimensional point cloud to obtain a sub-pixel aligned three-dimensional point cloud; inputting the polarized high dynamic range surface image and the sub-pixel aligned three-dimensional point cloud into a cross-modal defect attribution network for defect attribution detection, obtaining a defect process cause-effect graph; performing electrical safety dead zone verification based on the macro-infrared dual-mode image to obtain a safety margin index; and performing reverse tracing of production defects based on the defect process cause-effect graph and the safety margin index to locate the defect source process node.

[0007] One or more technical solutions provided in this application have at least the following beneficial effects:

[0008] Synchronous multimodal imaging at a composite inspection station, combining polarized high-dynamic surface images, laser-scanned 3D point clouds, and macro-infrared dual-mode images, enables comprehensive acquisition of surface features, 3D structure, and electrical safety information of copper busbars, providing rich multidimensional data for subsequent analysis. An iterative nearest-point algorithm precisely registers the laser-scanned 3D point cloud with the CAD model, resulting in a sub-pixel aligned 3D point cloud. This algorithm eliminates scanning errors and positioning deviations, accurately aligning the 3D point cloud with the CAD model and ensuring the accuracy of subsequent defect analysis. Finally, combining polarized high-dynamic surface images with the sub-pixel aligned 3D point cloud, a cross-modal defect attribution network is used for defect detection. The system generates a defect cause-and-effect diagram, which visually displays the causes of defects on the copper busbar surface. It provides specific process nodes for the source of each defect in production, thus providing a precise basis for process improvement and quality control. Based on micro-infrared dual-mode images, it performs electrical safety dead zone verification and obtains a safety margin index, which can identify potential electrical safety hazards in advance and make timely process adjustments or repairs, thereby improving the safety of power copper busbars. According to the defect cause-and-effect diagram and the safety margin index, it performs reverse tracing of production defects, which can locate defects to specific production links or processes. This provides a basis for optimizing the production process, reduces the occurrence of subsequent quality problems, and improves the controllability and reliability of the production process.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of a visual inspection system for the production quality of power copper busbars, provided as an embodiment of this application.

[0011] Figure 2 This is a schematic flowchart of a visual inspection method for the production quality of power copper busbars provided in an embodiment of this application.

[0012] Figure labeling: Multimodal imaging module 10, spatial registration module 20, defect attribution detection module 30, dead zone verification module 40, reverse tracing module 50. Detailed Implementation

[0013] 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.

[0014] Example 1, as Figure 1 As shown in the figure, this application embodiment provides a visual inspection system for the production quality of power copper busbars, the system comprising: The multimodal imaging module 10 drives a composite inspection station deployed at the end of the production line to perform synchronous multimodal imaging on the copper busbars to be inspected conveyed by the conveyor belt, obtaining polarized high dynamic surface images, laser-scanned 3D point clouds, and macro-infrared dual-mode images; the spatial registration module 20 is used to construct a spatial transformation matrix based on the center feature points of the mounting holes of the copper busbars to be inspected through an iterative nearest-point algorithm, and to perform CAD model spatial registration on the laser-scanned 3D point cloud to obtain subpixel-aligned 3D point clouds; the defect attribution detection module 30 is used to input the polarized high dynamic surface images and subpixel-aligned 3D point clouds into a cross-modal defect attribution network for defect attribution detection, obtaining a defect process cause-effect graph; the dead zone verification module 40 is used to perform electrical safety dead zone verification based on the macro-infrared dual-mode images to obtain a safety margin index; and the reverse tracing module 50 is used to perform reverse tracing of production defects based on the defect process cause-effect graph and the safety margin index to locate the defect source process node.

[0015] Furthermore, the defect attribution detection module 30 is used to perform the following operation steps: The polarized high dynamic surface image is input into a 2D defect segmentation model for surface defect type identification, and pixel-level defect segmentation maps are marked. The subpixel-aligned 3D point cloud is input into a 3D geometric analysis model for geometric defect feature identification, and 3D defect topological features are output. The pixel-level defect segmentation map is projected onto a triangular mesh surface reconstructed from the subpixel-aligned 3D point cloud, and cross-modal feature alignment is performed with the 3D defect topological features to obtain a multimodal defect feature map. After obtaining the process parameter records of the power copper busbar to be inspected from the MES module, the process parameter records and the multimodal defect feature map are spliced ​​in the input layer of the attribution network, and defect attribution detection is performed through the cross-modal defect attribution network to generate the defect process cause-effect graph.

[0016] Furthermore, the defect attribution detection module 30 is used to perform the following operation steps: After performing preprocessing including dynamic range correction and polarization difference fusion on the polarized high dynamic surface image, multi-scale features are extracted by the residual convolutional network of the 2D defect segmentation model. Defect response enhancement based on a channel-space dual-path attention mechanism is applied to the multi-scale features to obtain an enhanced feature map. The enhanced feature map is input into the decoder path of the 2D defect segmentation model, and multi-scale feature fusion is performed through a multi-receptive field feature fusion module, followed by progressive upsampling to output a defect probability map. The edge accuracy of the defect probability map is optimized using a conditional random field to generate the pixel-level defect segmentation map.

[0017] Furthermore, the defect attribution detection module 30 is used to perform the following operation steps: Based on a preset statistical threshold, the statistical thresholds of the subpixel aligned 3D point cloud are traversed to filter out outliers, resulting in a clean 3D point cloud. Geometric defect topology correlation analysis is performed on the clean 3D point cloud to obtain defect topology correlation features. Thin plate spline deformation analysis is performed on the flat areas of the clean 3D point cloud to obtain global deformation features. The defect topology correlation features and global deformation features are fused to output the 3D defect topology features.

[0018] Furthermore, the defect attribution detection module 30 is used to perform the following operation steps: The clean 3D point cloud is subjected to defect clustering segmentation based on Gaussian curvature calculation to obtain a set of geometric defect regions; connected component analysis is used to describe the defect features of the set of geometric defect regions to obtain a marked defect structure; a defect adjacency matrix is ​​established based on the marked defect structure, and the defect topological attributes are calculated to obtain the defect topological association features.

[0019] Furthermore, the defect attribution detection module 30 is used to perform the following operation steps: The multimodal defect feature map is rasterized at a preset spatial resolution to generate a spatial feature matrix; the process parameter records are vectorized into process parameter vectors; the spatial feature matrix is ​​flattened into feature vectors and process parameter vectors at the input layer of the attribution network, and then concatenated with them. After dimensionality reduction by a fully connected layer, defect fusion features are obtained; after constructing heterogeneous graph features based on the defect fusion features, the node features are updated in the cross-modal defect attribution network via a multi-layer graph convolutional network, the causal strength from process parameter nodes to defect feature nodes is calculated, and the defect process causal graph is pruned and output.

[0020] Furthermore, based on the defect fusion features, process nodes and defect nodes are associated to construct the heterogeneous graph features that include process parameter nodes, defect feature nodes, and process defect influence edges.

[0021] Furthermore, the multimodal imaging module 10 is used to perform the following operational steps: When the conveyor belt optical sensor detects that the copper busbar to be inspected has entered the composite inspection station, it triggers the PLC to control the servo motor to drive the linear slide to move at a constant speed. During the constant speed movement of the linear slide, a coaxial polarized light source is used in conjunction with an area array camera to perform multi-exposure frame acquisition, and a polarized high dynamic surface image that suppresses coating reflection is generated through a fusion algorithm. During the constant speed movement of the linear slide, a blue laser projector is used in conjunction with a high-speed camera to synchronously generate the laser scanning three-dimensional point cloud based on triangulation and bind encoder position tags. When the linear slide moves to the preset key area, the telecentric macro lens and short-wave infrared camera are activated to acquire images in different modes to obtain the macro infrared dual-mode image.

[0022] Furthermore, the spatial registration module 20 is used to perform the following operation steps: Curvature clustering segmentation is performed on the laser-scanned 3D point cloud to extract multiple mounting hole point cloud subsets. Then, a cylindrical model is fitted using the RANSAC algorithm to calculate the 3D coordinates of the centers of the multiple mounting holes. The reference center coordinates of the multiple mounting holes are obtained by parsing the STEP file of the copper busbar to be inspected. A descriptor is constructed based on the mounting hole topology. After topological constraint matching of the 3D coordinates of the centers of the multiple mounting holes and the reference center coordinates of the multiple mounting holes, the rigid transformation is calculated by singular value decomposition to obtain the initial spatial transformation matrix. Using the initial spatial transformation matrix as the initial value, multi-scale iterative nearest-point optimization is performed to obtain the sub-pixel aligned 3D point cloud.

[0023] Through the detailed description of a visual inspection method for the production quality of power copper busbars in this specification, those skilled in the art will clearly understand the visual inspection system for the production quality of power copper busbars in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to in the method section.

[0024] Example 2, based on the same inventive concept as the visual inspection system for the production quality of power copper busbars in the foregoing examples, such as... Figure 2 As shown in the figure, this application provides a visual inspection method for the production quality of power copper busbars, the method comprising: A100: Drives the composite inspection station deployed at the end of the production line to perform synchronous multimodal imaging on the copper busbars to be inspected conveyed by the conveyor belt, and obtain polarized high dynamic surface images, laser scanning three-dimensional point clouds and macro infrared dual-mode images.

[0025] By setting up a composite inspection station at the end of the production line, synchronous multimodal imaging is performed on the copper busbars to be inspected conveyed on the conveyor belt. This multimodal imaging includes the acquisition of the following image data: polarized high dynamic surface images, using a coaxial polarized light source and an area array camera to acquire images in multiple exposure frames. This, combined with polarized light, eliminates reflected light from the copper busbar surface, making surface defects such as scratches and dents more obvious in the images. Through fusion algorithms, these images can effectively suppress surface coating reflections; laser scanning 3D point cloud, using a blue laser projector and a high-speed camera to generate 3D point cloud data of the copper busbar during simultaneous triangulation. By combining encoder position tags, the acquired point cloud data is ensured to have accurate spatial coordinate information, facilitating subsequent spatial registration and defect analysis; and macro infrared dual-mode images, using a telecentric macro lens and a short-wave infrared camera, acquiring images of the copper busbar based on two different modes, including visible light and infrared, to obtain infrared imaging of its surface and interior. This helps detect electrical safety issues in the copper busbar, such as dead zones and other defects affecting electrical performance.

[0026] A200: Based on the center feature point of the mounting hole of the copper busbar to be detected, a spatial transformation matrix is ​​constructed by using an iterative nearest point algorithm. The laser-scanned three-dimensional point cloud is then spatially registered with the CAD model to obtain a sub-pixel aligned three-dimensional point cloud.

[0027] The iterative nearest-point algorithm calculates a spatial transformation matrix based on the center feature points of the mounting holes in the copper busbar to be inspected. The center points of the mounting holes are deterministic within the geometric model of the copper busbar, serving as the fundamental features for alignment. The iterative nearest-point algorithm is then used to register the laser-scanned 3D point cloud with the CAD model space of the copper busbar. This algorithm iteratively optimizes the matching degree between the point cloud data and the model to obtain a high-precision spatial transformation matrix. After optimization, the resulting 3D point cloud data undergoes sub-pixel alignment for even finer spatial matching. Sub-pixel alignment refers to improving the point cloud's precision to less than one pixel level, ensuring high-precision 3D reconstruction and subsequent defect analysis.

[0028] A300: Input the polarized high dynamic surface image and subpixel aligned 3D point cloud into the cross-modal defect attribution network for defect attribution detection to obtain a defect process cause-effect graph.

[0029] Polarized high-dynamic surface images and subpixel-aligned 3D point clouds are input into a cross-modal defect attribution network. This network integrates data from two different sources to perform defect detection and attribution analysis. In this network, surface defect types are first identified using a 2D defect segmentation model. This process involves pixel-level analysis of the polarized high-dynamic surface images to mark defects such as scratches and bulges on the surface of the power busbar and generate a defect segmentation map. For the 3D point cloud data, a 3D geometric analysis model is used for geometric defect identification. The model extracts geometric features from the point cloud to identify geometric defects on the busbar surface, such as deformation and cracks. These features provide more intuitive defect information and, together with surface defects, provide data support for subsequent analysis. The defect features extracted from the 2D defect segmentation model and the 3D geometric analysis model are fused. By aligning the pixel-level information of surface defects with the topological features of 3D defects, a multimodal defect feature map is generated. This image integrates defect information from different perspectives and can comprehensively reflect the production quality issues of the busbar.

[0030] Further, process parameter records for the copper busbars to be inspected, such as stamping pressure and electroplating temperature, are obtained from the MES (Manufacturing Execution System) module. These process parameters are then stitched together with the defect feature map and fused through a cross-modal defect attribution network to generate a defect process cause-effect graph. This graph clearly indicates the source of each defect. For example, the network might indicate: "Surface scratches are caused by die wear in the stamping workshop" or "Plating bubbles are caused by excessive electroplating temperature." This cause-effect graph not only displays the defect itself but also identifies the process steps that lead to the defect through correlation analysis, providing precise information for subsequent quality control and improvement.

[0031] A400: Based on the aforementioned micro-infrared dual-mode image, electrical safety dead zone verification is performed to obtain the safety margin index.

[0032] Micro-infrared dual-mode imaging, which combines infrared and visible light images, is used to image power busbars. Infrared images provide details related to electrical safety, such as the heat distribution of current flow and potential localized overheating areas. Electrical dead zones refer to areas where current cannot flow effectively during electrical connections. These areas can lead to overheating or malfunctions in the power busbar during use. Analysis of micro-infrared dual-mode images verifies these potential electrical dead zones. Based on the image analysis results, a safety margin index is calculated. This index represents the "margin" or "safety margin" of the power busbar in actual operation. A low index indicates the presence of electrical dead zones, which may affect the reliability and safety of the power busbar; a high index indicates good electrical safety.

[0033] A500: Based on the aforementioned defect process cause-effect diagram and safety margin index, reverse tracing of production defects is performed to locate the defect source process node.

[0034] By combining defect cause-and-effect diagrams with safety margin indices for reverse tracing, the relationship between different process parameters and defects is analyzed to pinpoint specific defect-originating process nodes. For example, die wear in a stamping process might be the source of scratches, while improper temperature settings in an electroplating process might be the root cause of plating bubbles. Through this reverse tracing, problems on the production line can be quickly located, allowing for targeted improvements.

[0035] Furthermore, the polarization high dynamic range surface image and subpixel aligned 3D point cloud are input into a cross-modal defect attribution network for defect attribution detection to obtain a defect process causal graph. The method includes: A310: Input the polarized high dynamic surface image into a 2D defect segmentation model for surface defect type identification and label pixel-level defect segmentation maps; A320: Input the subpixel-aligned 3D point cloud into a 3D geometric analysis model for geometric defect feature identification and output 3D defect topological features; A330: Project the pixel-level defect segmentation map onto a triangular mesh surface reconstructed from the subpixel-aligned 3D point cloud, and perform cross-modal feature alignment with the 3D defect topological features to obtain a multimodal defect feature map; A340: After obtaining the process parameter records of the power copper busbar to be inspected from the MES module, splice the process parameter records and the multimodal defect feature map at the attribution network input layer, and perform defect attribution detection through the cross-modal defect attribution network to generate the defect process cause-effect graph.

[0036] Polarized high-dynamic surface images are input into a 2D defect segmentation model. These images, processed with polarized light to suppress coating reflection, make surface defects on the copper busbar more apparent, such as scratches, bulges, and stains. The 2D defect segmentation model, built on deep learning technologies like convolutional neural networks, performs pixel-level classification and segmentation of surface defects. After training, the model can identify the specific location and type of surface defects. During defect type identification, the model automatically identifies surface defects in the image and labels them by type, such as scratches, bulges, and bubbles, segmenting them into independent regions. This process provides detailed defect information, supporting subsequent quality control and process optimization. Finally, a pixel-level defect segmentation map is output, where each defect region is accurately labeled and segmented. This image provides detailed defect information, precisely locating quality problems on the surface of the power copper busbar.

[0037] Subpixel-aligned 3D point clouds are input into a 3D geometric analysis model, which is mainly used for in-depth analysis of the geometry of the copper busbar to identify existing geometric defects, such as surface deformation, cracks, and uneven thickness. The 3D geometric analysis model identifies the topological features of these geometric defects and outputs them as 3D defect topological features. These topological features reflect the specific morphology of the defects, such as their location, shape, and relationship with other parts of the copper busbar.

[0038] The pixel-level defect segmentation map is combined with the sub-pixel aligned 3D point cloud. Through projection, the 2D image data is mapped onto the point cloud in 3D space. In this step, the 3D point cloud is used to reconstruct the surface model of the copper busbar, represented by a triangular mesh surface. This triangular mesh meticulously reconstructs the surface of the copper busbar, ensuring that the 3D position of each defect accurately corresponds to the defect information in the image. The defect features extracted from the pixel-level defect segmentation map are aligned with the geometric defect features extracted from the 3D point cloud. This process includes registration and fusion of features from different modal data. By combining the 2D and 3D information of the defects through cross-modal feature alignment, a more comprehensive multimodal defect feature map is generated. This map contains combined information on surface defects (from the 2D image) and geometric defects (from the 3D point cloud), comprehensively displaying all defect features present in the copper busbar.

[0039] The process parameter records related to the copper busbar to be inspected are obtained from the MES module (Manufacturing Execution System). These process parameters include specific process information during the copper busbar production process, such as the pressure, temperature settings, and time parameters used during stamping. This information can provide background information that may lead to defects during the production process.

[0040] Process parameter records and multimodal defect feature maps are stitched together to form comprehensive input data, which includes copper busbar production information and all detected defect features. This stitched data is then input into a cross-modal defect attribution network, which performs defect attribution analysis by analyzing the relationship between the spatial characteristics of defects and process parameters. After processing by the attribution network, a defect process cause-and-effect graph is generated. This graph clearly identifies the source and responsibility for each defect, such as: "Surface scratches are caused by die wear in the stamping workshop," and "Bubbles in the electroplating layer are caused by excessively high electroplating solution temperature." These causal relationships help trace back to specific production stages, providing a clear basis for quality control in production.

[0041] Furthermore, the method involves inputting the polarized high dynamic range surface image into a 2D defect segmentation model for surface defect type identification and labeling pixel-level defect segmentation maps. A311: After performing preprocessing including dynamic range correction and polarization difference fusion on the polarization high dynamic surface image, the image is input into the residual convolutional network of the 2D defect segmentation model to extract multi-scale features; A312: The multi-scale features are enhanced with defect response based on a channel-space dual-path attention mechanism to obtain an enhanced feature map; A313: The enhanced feature map is input into the decoder path of the 2D defect segmentation model, and multi-scale feature fusion is performed through a multi-receptive field feature fusion module, and progressive upsampling is performed to output a defect probability map; A314: The edge accuracy of the defect probability map is optimized through a conditional random field to generate the pixel-level defect segmentation map.

[0042] High dynamic range (HMR) polarized surface images often contain bright or dark areas, which can affect subsequent analysis. Dynamic range correction adjusts the brightness and contrast of the image to make the information in the image more balanced, thereby reducing the impact of reflected light or low-contrast areas on defect identification. When acquiring polarized images, due to surface reflection, polarization differential fusion aims to combine images from different polarization angles. By removing or reducing the influence of reflection, surface defects become more apparent, and the fused image can better display the details of the copper busbar surface, such as scratches and bulges.

[0043] The preprocessed image is input into the residual convolutional network of the 2D defect segmentation model. The residual convolutional network avoids the gradient vanishing problem through residual blocks and can extract multi-scale features from the image. Through this network, low-level and high-level features in the image can be extracted, preparing for defect detection.

[0044] A dual-path channel-spatial attention mechanism is used to enhance the response of defect regions in the image: the channel attention mechanism weights the importance of each channel (i.e., different feature dimensions of the image), emphasizing those features that are more important for defect identification; the spatial attention mechanism strengthens the features of the defect region and weakens the features of the background region in the spatial dimension of the image, ensuring that the model focuses more on the defect. By fusing the above two attention mechanisms, an enhanced feature map is obtained. The enhanced feature map is characterized by a significant enhancement of the response of the defect region, while background noise and irrelevant parts are suppressed, thereby improving the accuracy of subsequent defect identification.

[0045] The enhanced feature map is input into the decoder path of the 2D defect segmentation model. The decoder restores the high-dimensional feature map to the original image size and performs defect prediction. The multi-receptive field feature fusion module in the decoder path uses convolutional operations with different receptive fields to fuse feature information at multiple scales. This allows the network to identify defects at different scales, such as handling small scratches or large bulges. Progressive upsampling is also performed in the decoder path, gradually increasing the resolution of the feature map to make the final defect probability map more detailed and accurate. Upsampling helps restore spatial details in the image, thereby enhancing the representation of defect edges. Finally, through the above processing, a defect probability map is output, showing the probability that each pixel in the image belongs to a defect region. This is a pixel-based image where the value of each pixel represents the probability that the location is a defect, expressed using color or grayscale values.

[0046] Conditional Random Fields (CRFs) are graphical models that optimize image segmentation by modeling dependencies between pixels. In this step, CRFs are used to optimize the edge accuracy of the defect probability map, reducing noise or discontinuous areas in the segmentation result, making the defect edges smoother and more accurate. After CRF optimization, the final pixel-level defect segmentation map contains accurate defect edges and location markings. In this map, the defect region is clearly marked with sharp edges, facilitating subsequent defect analysis and processing.

[0047] Furthermore, the method involves inputting the subpixel-aligned 3D point cloud into a 3D geometric analysis model, identifying geometric defect features, and outputting 3D defect topological features. A321: Based on a preset statistical threshold, traverse the statistical threshold of the sub-pixel aligned 3D point cloud to filter out outliers and obtain a clean 3D point cloud; A322: Perform geometric defect topology correlation analysis on the clean 3D point cloud to obtain defect topology correlation features; A323: Perform thin plate spline deformation analysis on the flat areas of the clean 3D point cloud to obtain global deformation features; A324: Fuse the defect topology correlation features and global deformation features to output the 3D defect topology features.

[0048] Based on the nature of the data, a preset statistical threshold is first set. This threshold, based on statistical characteristics such as the distribution, density, or standard deviation of the point cloud, is used to distinguish normal data points from potential outliers. The sub-pixel aligned 3D point cloud is traversed. For each point, its relationship with other points is calculated, such as the distance between points or differences in certain geometric attributes. Then, the preset statistical threshold is used to determine whether the point is an outlier. If a point's attributes exceed the set threshold range, it is identified as an outlier and automatically removed from the 3D point cloud, retaining only valid point data. After outlier filtering, the resulting point cloud data is cleaner and more coherent, termed a clean 3D point cloud. This point cloud data can more accurately reflect the true geometric shape of the copper busbar, providing a foundation for subsequent defect analysis.

[0049] After obtaining a clean 3D point cloud, topological analysis is performed on the point cloud data using geometric analysis techniques. The goal is to identify geometric defects present on the surface or structure of the copper busbar, such as: shape variations, including bending, twisting, or irregular deformation of the copper busbar surface; and geometric damage, such as cracks and defects. Through topological correlation analysis, the topological features of these geometric defects are extracted. These features describe the location, shape, and relative relationship of the defects to other parts in 3D space.

[0050] Flat regions in the clean 3D point cloud are identified. These flat regions correspond to the surface of the copper busbar or parts without significant geometric defects. For these flat regions, the thin plate spline deformation analysis method is used. The thin plate spline is a mathematical tool that can perform high-precision deformation fitting on planar regions. This method analyzes the deformation of the copper busbar surface, such as warping and bending, by minimizing the deformation energy, and outputs the global deformation characteristics of the region, that is, the large-scale deformation of the copper busbar surface. The global deformation characteristics can provide the overall deformation of the copper busbar during the production process, which helps to further judge its performance and usage status.

[0051] The obtained defect topology correlation features are fused with global deformation features to combine local defects and overall deformation, comprehensively reflecting the geometric quality of the copper busbar. After feature fusion, a three-dimensional defect topology feature is output, which can comprehensively describe the geometric defects and deformation of the copper busbar, providing rich information for subsequent defect analysis, process optimization, and quality control.

[0052] Furthermore, geometric defect topology correlation analysis is performed on the clean 3D point cloud to obtain defect topology correlation features. The method includes: A3221: Perform defect clustering segmentation on the clean 3D point cloud based on Gaussian curvature calculation to obtain a set of geometric defect regions; A3222: Use connected component analysis to describe the defect features of the set of geometric defect regions to obtain a marked defect structure; A3223: Establish a defect adjacency matrix based on the marked defect structure and perform defect topology attribute calculation to obtain the defect topology association features.

[0053] Gaussian curvature is a geometric measure describing the local morphology of a surface, reflecting the degree of curvature of the surface at a given point. By calculating the Gaussian curvature of each point in a point cloud, areas of surface deformation, such as depressions and bulges, can be effectively identified. Gaussian curvature calculation is based on the local neighborhood of the point cloud, calculating the curvature value of each point by fitting the surface. Regions with higher curvature indicate surface defects or irregular deformations. Based on the calculated Gaussian curvature values, the points in the point cloud are clustered according to their curvature. High curvature regions correspond to geometric defects, such as depressions, bulges, and cracks. Using clustering algorithms such as k-means and DBSCAN, points with similar curvature values ​​are grouped into the same category, forming a set of geometric defect regions. Each defect region represents a local deformation or irregular area in the point cloud, providing a foundation for subsequent defect analysis.

[0054] Connected component analysis refers to the analysis of point cloud data to identify connected sets of points. This analysis determines which points belong to the same geometric defect. For example, if multiple sets of points are connected to the edge of a defect through some adjacent points, these sets can be considered a connected component. For each connected component, its shape features, such as area, perimeter, and shape complexity, as well as spatial features, such as relative position and size, are extracted. These features describe the nature and extent of the defect. Furthermore, the defect type is labeled based on its shape characteristics, such as depressions, bulges, and cracks. After connected component analysis, each defect region is labeled as a marked defect structure. Each marked defect structure contains a description of the defect features in that region, facilitating subsequent processing and analysis.

[0055] A defect adjacency matrix is ​​a matrix that describes the spatial relationships between defects. Each element in the matrix represents the adjacency relationship between two defect regions. For example, if two defect regions are spatially adjacent (i.e., their boundaries overlap or touch), the corresponding element in the adjacency matrix is ​​1; if they are not adjacent, it is 0. This adjacency matrix provides a visual representation of the connectivity between defects, aiding in further analysis of defect propagation and impact. Based on the adjacency matrix, the topological properties of defects are calculated, such as connectivity between defects, defect distribution patterns, and the morphology and layout of defect regions. These topological properties describe the interrelationships of defects in three-dimensional space, providing detailed information for defect source tracing and quality control. Finally, through the calculation of the adjacency matrix and topological properties, defect topological association characteristics are obtained. These characteristics reveal the spatial relationships and topological structure between defect regions, contributing to further analysis of defect root causes, propagation paths, and influencing factors.

[0056] Furthermore, after obtaining the process parameter records of the copper busbar to be inspected from the MES module, the process parameter records and the multimodal defect feature map are concatenated at the input layer of the attribution network, and defect attribution detection is performed through the cross-modal defect attribution network to generate the defect process cause-effect graph. The method includes: A341: Rasterize the multimodal defect feature map according to a preset spatial resolution to generate a spatial feature matrix; A342: Vectorize the process parameter records into process parameter vectors; A343: Flatten the spatial feature matrix into feature vectors and process parameter vectors in the input layer of the attribution network, concatenate them, and reduce the dimensionality through a fully connected layer to obtain defect fusion features; A344: After constructing heterogeneous graph features based on the defect fusion features, update the node features in the cross-modal defect attribution network through a multi-layer graph convolutional network, calculate the causal strength from process parameter nodes to defect feature nodes, and prune and output the defect process causal graph.

[0057] The multimodal defect feature map is rasterized according to a preset spatial resolution. This means dividing the image into regular grid cells, each representing a certain spatial range. This ensures data uniformity and standardization, facilitating subsequent calculations and feature extraction. The rasterized image data generates a spatial feature matrix, where each element represents the feature information of the corresponding raster.

[0058] Transforming process parameter records into process parameter vectors involves converting discrete parameter values ​​into processable numerical vectors, enabling process parameters to be input into the attribution network along with other features for comprehensive analysis.

[0059] The obtained spatial feature matrix is ​​flattened into a one-dimensional feature vector. The flattening process transforms two-dimensional spatial information into one-dimensional data, facilitating concatenation and processing with other vectors. The flattened spatial feature vector is concatenated with the process parameter vector, and then dimensionality reduction is performed through a fully connected layer to remove redundant information and reduce the dimensionality of the data. The dimensionality-reduced feature vector is easier for subsequent network processing to form defect fusion features. The defect fusion features contain the spatial information of the defects and their related production process parameters.

[0060] Based on defect fusion features, a heterogeneous graph feature is constructed. This heterogeneous graph feature includes process parameter nodes, defect feature nodes, and the relationships between them. Nodes in the heterogeneous graph represent different entities, such as process parameters and defect features, and edges represent the relationships between them. The heterogeneous graph feature is input into a cross-modal defect attribution network. This network performs defect attribution by analyzing the relationships between process parameter nodes and defect feature nodes. By calculating the causal relationships between these nodes, the network reveals which process parameters are related to specific defects.

[0061] The cross-modal defect attribution network uses a graph convolutional network. By iteratively updating node features and leveraging adjacency relationships in the graph structure, it enhances the information of each node. Through multiple convolutional operations, the network can learn more complex inter-node relationships, thus enabling more accurate defect attribution analysis. The graph convolutional network calculates the causal strength between process parameter nodes and defect feature nodes, i.e., the degree to which each process parameter affects a specific defect. This helps reveal which production stages or process steps are the main causes of defects. After calculating the causal strength, the graph is pruned to remove unimportant nodes and edges. Finally, a defect process causal graph is output, clearly indicating the root process step of each defect. For example, the graph might show "Defect A is caused by excessive pressure in the stamping process" or "Defect B is caused by excessive electroplating temperature."

[0062] Furthermore, based on the defect fusion features, process nodes and defect nodes are associated to construct the heterogeneous graph features that include process parameter nodes, defect feature nodes, and process defect influence edges.

[0063] In the preceding steps, defect fusion features were obtained by fusing spatial feature matrices and process parameter vectors, including spatial information about defects and their relationship with production process parameters. Based on these defect fusion features, the association between process parameter nodes and defect feature nodes was established by analyzing the relationship between defect features and process parameters. Process parameter nodes represent various parameters in the production process, such as stamping pressure and electroplating temperature; defect feature nodes represent detected defect types, such as surface scratches, bulges, and bubbles. A heterogeneous graph feature structure was constructed, containing process parameter nodes, defect feature nodes, and process defect influence edges. Process defect influence edges represent the causal or influence relationship between process parameters and defects; for example, parameter settings in a certain process step, such as temperature and pressure, directly affect the occurrence of a certain defect.

[0064] Furthermore, the method involves driving a composite inspection station deployed at the end of the production line to simultaneously perform multimodal imaging on the copper busbars to be inspected conveyed by the conveyor belt, obtaining polarized high dynamic surface images, laser-scanned three-dimensional point clouds, and macro-infrared dual-mode images. A110: When the conveyor belt optical sensor detects that the copper busbar to be tested has entered the composite detection station, it triggers the PLC to control the servo motor to drive the linear slide to move at a constant speed; A120: During the constant speed movement of the linear slide, a coaxial polarized light source is used in conjunction with an area array camera to perform multi-exposure frame acquisition, and the polarized high dynamic surface image that suppresses coating reflection is generated through a fusion algorithm; A130: During the constant speed movement of the linear slide, a blue laser projector is used in conjunction with a high-speed camera to synchronously generate the laser scanning three-dimensional point cloud based on triangulation and bind the encoder position tag; A141: When the linear slide moves to the preset key area, the telecentric macro lens and short-wave infrared camera are activated to acquire images in different modes to obtain the macro infrared dual-mode image.

[0065] The conveyor belt optical sensor is used to detect whether the copper busbar to be inspected has entered the composite inspection station. When the copper busbar on the conveyor belt moves to the designated position, the optical sensor detects its presence and sends a signal. After receiving the signal from the conveyor belt optical sensor, the PLC (Programmable Logic Controller) triggers the servo motor. The task of the servo motor controlled by the PLC is to drive the linear slide to move at a constant speed along the conveyor belt. The linear slide moves the copper busbar to be inspected to the inspection station at a constant speed. This process ensures that the copper busbar stays at the composite inspection station for a sufficient time to perform various inspection tasks, such as polarization high dynamic surface image acquisition and 3D point cloud scanning.

[0066] During the uniform movement of the linear slide, image acquisition is performed using a combination of a coaxial polarized light source and an area array camera. The polarized light source reduces the influence of reflected light, especially the reflection from the copper busbar surface coating, by altering the direction of light vibration. The area array camera, in conjunction with the polarized light source, provides a wider field of view and further suppresses reflected light, making defects on the copper busbar surface more clearly visible. To further enhance the dynamic range of the image, a multi-exposure frame acquisition technique is employed. By capturing multiple frames at different exposure times, areas of varying brightness can be covered, resulting in richer details. A fusion algorithm synthesizes these multiple exposure images to obtain a polarized high dynamic range surface image. This image suppresses reflection from the coating and clearly displays surface defects.

[0067] The linear slide continues its uniform movement, and a blue laser projector and a high-speed camera work synchronously. The blue laser projector emits laser lines that illuminate the surface of the copper busbar, while the high-speed camera simultaneously captures the shape formed by the laser lines on the copper busbar surface. The three-dimensional shape of the surface is then calculated based on the deformation of the laser lines. Using the principle of triangulation, the spatial coordinates of each point are calculated based on the angular difference between the laser lines and the camera. This method accurately generates three-dimensional point cloud data of the copper busbar. During this process, encoder position tags are also bound to the three-dimensional point cloud data. The encoder records the position of the slide, ensuring a one-to-one correspondence between the generated three-dimensional point cloud data and the precise position of the power copper busbar on the conveyor belt. Through these steps, a high-precision laser-scanned three-dimensional point cloud is generated. This data reflects the three-dimensional structure of the copper busbar surface and is used for subsequent defect analysis and location.

[0068] When the linear slide moves to a preset key position, image acquisition is performed. The preset key area is determined based on specific regions of the power busbar, such as areas with electrical dead zones or surface details. The telecentric macro lens is activated, possessing high magnification and high-precision imaging capabilities, clearly capturing the fine structure of the power busbar surface, especially areas with defects or minor electrical problems. Simultaneously, a short-wave infrared camera also begins operation, acquiring infrared images of the power busbar surface and its interior. The infrared camera can detect heat distribution, especially areas that may cause safety issues, such as electrical dead zones. These areas are typically undetectable by visible light images, but infrared imaging can identify potential electrical problems. Combining the acquisitions from the telecentric macro lens and the short-wave infrared camera yields a macro infrared dual-mode image. This image provides detailed information about the power busbar and electrical safety, providing crucial data for subsequent safety analysis and defect location.

[0069] Furthermore, based on the center feature points of the mounting holes of the copper busbar to be detected, a spatial transformation matrix is ​​constructed using an iterative nearest-point algorithm. This matrix is ​​then used to perform CAD model spatial registration on the laser-scanned 3D point cloud, resulting in a sub-pixel aligned 3D point cloud. The method includes: A210: Perform curvature clustering segmentation on the laser-scanned 3D point cloud, extract multiple mounting hole point cloud subsets, fit a cylindrical model using the RANSAC algorithm, and calculate the 3D coordinates of the centers of multiple mounting holes; A220: Parse the STEP file of the copper busbar to be inspected to obtain the reference center coordinates of multiple mounting holes; A230: Construct a descriptor based on the mounting hole topology, perform topological constraint matching between the 3D coordinates of the centers of multiple mounting holes and the reference center coordinates of multiple mounting holes, and calculate the rigid transformation through singular value decomposition to obtain the initial spatial transformation matrix; A240: Use the initial spatial transformation matrix as the initial value, perform multi-scale iterative nearest-point optimization to obtain the sub-pixel aligned 3D point cloud.

[0070] By calculating the curvature of each point in the laser-scanned 3D point cloud, different regions on the copper busbar surface are identified. Mounting holes are located on the copper busbar surface; regions with lower curvature represent flat areas, while the edges of the mounting holes exhibit greater curvature variations. Based on the calculated curvature information, curvature clustering segmentation is performed, meaning the points in the point cloud are divided into different clusters with similar curvature values. This method allows for the separation of a subset of the point cloud representing the mounting hole region from the total point cloud. On the extracted mounting hole point cloud subset, the RANSAC (Random Sample Consensus) algorithm is used for cylindrical model fitting. RANSAC effectively identifies the geometric model in the point cloud and removes outliers. By fitting the cylindrical model, the shape and location of the mounting holes can be accurately described. After the cylindrical model fitting is complete, the 3D coordinates of the center of each mounting hole are calculated. These center coordinates will be used as key features in the subsequent registration step for alignment with the mounting hole positions in the CAD model.

[0071] STEP files are a type of CAD model file format that contains detailed geometric information about the power copper busbars. In this step, the file is parsed to extract the mounting hole location data and obtain the coordinates of the reference centers of multiple mounting holes. These coordinates come from the design drawings of the copper busbars and are the ideal positions in the CAD model, which are used for subsequent spatial registration.

[0072] By modeling the topology of each mounting hole, a descriptor is constructed, representing the feature information of the relative position and arrangement of the mounting holes. This descriptor characterizes the layout and interrelationships of the mounting holes, aiding in matching between the point cloud and the CAD model. Based on the descriptor, topological constraint matching is performed between the center coordinates of the mounting holes obtained from laser scanning and the reference center coordinates of the mounting holes extracted from the CAD model. This matching ensures the accuracy of registration by constraining the relative positional relationship between the mounting holes in the point cloud and the mounting holes in the CAD model. During the matching process, an initial spatial transformation matrix is ​​calculated using singular value decomposition. This transformation matrix aligns the coordinates of the mounting holes in the laser-scanned point cloud with the reference center coordinates in the CAD model, providing preliminary registration results for subsequent optimization steps.

[0073] The spatial registration is further optimized using a multi-scale iterative nearest-neighbor algorithm. This algorithm iteratively minimizes the distance error between the laser-scanned point cloud and the CAD model, gradually improving registration accuracy. The multi-scale approach allows optimization at different scales, ensuring good alignment from large-scale to detailed levels. After multi-scale iterative nearest-neighbor optimization, the resulting sub-pixel aligned 3D point cloud has higher accuracy. At this stage, each point in the point cloud is aligned to the sub-pixel level, ensuring accurate spatial registration with the CAD model.

[0074] 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 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 visual inspection system for the production quality of power copper busbars, characterized in that, The system includes: The multimodal imaging module is used to drive the composite inspection station deployed at the end of the production line to perform synchronous multimodal imaging of the copper busbars to be inspected conveyed by the conveyor belt, and obtain polarized high dynamic surface images, laser scanning three-dimensional point clouds and macro infrared dual-mode images. The spatial registration module is used to construct a spatial transformation matrix based on the center feature points of the mounting holes of the copper busbar to be detected through an iterative nearest point algorithm, and to perform CAD model spatial registration on the laser-scanned three-dimensional point cloud to obtain a sub-pixel aligned three-dimensional point cloud. The defect attribution detection module is used to input the polarized high dynamic surface image and sub-pixel aligned 3D point cloud into the cross-modal defect attribution network for defect attribution detection, and obtain a defect process cause-effect graph. The dead zone verification module is used to perform electrical safety dead zone verification based on the macro infrared dual-mode image to obtain a safety margin index. The reverse tracing module is used to perform reverse tracing of production defects based on the defect process cause-effect graph and safety margin index, and to locate the defect source process node.

2. The visual inspection system for the production quality of power copper busbars as described in claim 1, characterized in that, The defect attribution detection module is used to perform the following steps: The polarized high dynamic surface image is input into a 2D defect segmentation model to identify the surface defect type and mark the pixel-level defect segmentation map. The subpixel aligned 3D point cloud is input into a 3D geometric analysis model to identify geometric defect features and output 3D defect topological features. The pixel-level defect segmentation map is projected onto the triangular mesh surface reconstructed from the sub-pixel aligned 3D point cloud, and cross-modal feature alignment is performed with the 3D defect topological features to obtain a multimodal defect feature map. After obtaining the process parameter records of the copper busbar to be inspected from the MES module, the process parameter records and the multimodal defect feature map are spliced ​​together in the input layer of the attribution network, and the defect attribution detection is performed through the cross-modal defect attribution network to generate the defect process cause-effect graph.

3. The visual inspection system for the production quality of power copper busbars as described in claim 2, characterized in that, The defect attribution detection module is used to perform the following steps: After performing preprocessing including dynamic range correction and polarization difference fusion on the polarized high dynamic surface image, the residual convolutional network of the 2D defect segmentation model is input to extract multi-scale features; The multi-scale features are enhanced with a defect response based on a channel-space dual-path attention mechanism to obtain an enhanced feature map; The enhanced feature map is input into the decoder path of the 2D defect segmentation model, multi-scale feature fusion is performed through the multi-receptive field feature fusion module, and progressive upsampling is performed to output a defect probability map. The edge accuracy of the defect probability map is optimized by using a conditional random field to generate the pixel-level defect segmentation map.

4. The visual inspection system for the production quality of power copper busbars as described in claim 3, characterized in that, The defect attribution detection module is used to perform the following steps: Based on a preset statistical threshold, the statistical thresholds of the sub-pixel aligned 3D point cloud are traversed to filter out outliers and obtain a clean 3D point cloud. Geometric defect topology correlation analysis was performed on the clean 3D point cloud to obtain defect topology correlation features; Thin-plate spline deformation analysis was performed on the flat area of ​​the clean 3D point cloud to obtain global deformation characteristics; The three-dimensional defect topology features are output by fusing the defect topology association features and global deformation features.

5. The visual inspection system for the production quality of power copper busbars as described in claim 4, characterized in that, The defect attribution detection module is used to perform the following steps: The clean 3D point cloud is subjected to defect clustering segmentation based on the Gaussian curvature calculation of the point cloud to obtain a set of geometric defect regions; Connectivity analysis is used to describe the defect features of the geometric defect region set, resulting in a marked defect structure. Based on the marked defect structure, a defect adjacency matrix is ​​established, and the defect topology attribute is calculated to obtain the defect topology association features.

6. The visual inspection system for the production quality of power copper busbars as described in claim 2, characterized in that, The defect attribution detection module is used to perform the following steps: The multimodal defect feature map is rasterized at a preset spatial resolution to generate a spatial feature matrix; The process parameter records are vectorized into process parameter vectors; In the input layer of the attribution network, the spatial feature matrix is ​​flattened into feature vectors and process parameter vectors, which are then concatenated and dimensionality reduced by a fully connected layer to obtain defect fusion features. After constructing heterogeneous graph features based on the defect fusion features, the node features are updated in the cross-modal defect attribution network via a multi-layer graph convolutional network. The causal strength from process parameter nodes to defect feature nodes is calculated, and the defect process causal graph is pruned and output.

7. The visual inspection system for the production quality of power copper busbars as described in claim 6, characterized in that, Based on the defect fusion features, process nodes and defect nodes are associated to construct the heterogeneous graph features that include process parameter nodes, defect feature nodes, and process defect influence edges.

8. The visual inspection system for the production quality of power copper busbars as described in claim 1, characterized in that, The multimodal imaging module is used to perform the following operation steps: When the light sensor on the conveyor belt detects that the copper busbar to be tested enters the composite detection station, it triggers the PLC to control the servo motor to drive the linear slide table to move at a constant speed. During the uniform movement of the linear slide, a coaxial polarized light source is used in conjunction with an area array camera to acquire multiple exposure frames, and a polarized high dynamic range surface image that suppresses coating reflection is generated through a fusion algorithm. During the uniform movement of the linear slide, a blue laser projector is used in conjunction with a high-speed camera to synchronously generate the laser-scanned three-dimensional point cloud based on triangulation and bind encoder position labels. When the linear slide moves to the preset key position, the telecentric macro lens and shortwave infrared camera are activated to acquire images in two modes, thus obtaining the macro infrared dual-mode image.

9. The visual inspection system for the production quality of power copper busbars as described in claim 1, characterized in that, The spatial registration module is used to perform the following operations: After performing curvature clustering segmentation on the laser-scanned 3D point cloud and extracting multiple mounting hole point cloud subsets, the cylindrical model is fitted using the RANSAC algorithm to calculate the 3D coordinates of the centers of multiple mounting holes. The coordinates of the reference centers of multiple mounting holes are obtained by parsing the STEP file of the copper busbar to be tested; Based on the topology of the mounting holes, a descriptor is constructed. After topological constraint matching of the three-dimensional coordinates of the centers of the multiple mounting holes and the coordinates of the reference centers of the multiple mounting holes, the initial spatial transformation matrix is ​​obtained by calculating the rigid transformation through singular value decomposition. Using the initial spatial transformation matrix as the initial value, multi-scale iterative nearest-point optimization is performed to obtain the sub-pixel aligned 3D point cloud.

10. A visual inspection method for the production quality of power copper busbars, characterized in that, Based on the implementation of the visual inspection system for production quality of power copper busbars according to any one of claims 1-9, the method includes: The composite inspection station deployed at the end of the production line drives synchronous multimodal imaging of the copper busbars to be inspected conveyed by the conveyor belt, and obtains polarized high dynamic surface images, laser scanning three-dimensional point clouds and macro infrared dual-mode images. Based on the center feature points of the mounting holes of the copper busbar to be detected, a spatial transformation matrix is ​​constructed by an iterative nearest point algorithm, and the laser-scanned three-dimensional point cloud is spatially registered with the CAD model to obtain a sub-pixel aligned three-dimensional point cloud. The polarized high dynamic surface image and subpixel aligned 3D point cloud are input into a cross-modal defect attribution network for defect attribution detection to obtain a defect process cause-effect graph. Electrical safety dead zone verification is performed based on the aforementioned micro-infrared dual-mode image to obtain a safety margin index; Based on the aforementioned defect process cause-effect graph and safety margin index, reverse tracing of production defects is performed to locate the defect source process node.

Images