Method for detecting three-dimensional topography defects of an electric cell

Abstract

The application discloses a battery cell three-dimensional morphology defect detection method, and belongs to the field of machine vision and industrial detection. Firstly, the ArUco mark feature is relied on to complete rough registration of a battery cell point cloud; then, plane fitting is carried out, and a constrained nonlinear minimization algorithm is combined to realize fine registration of the point cloud; subsequently, the curvature is used to distinguish feature grids and plane grids, and a differentiated strategy is used to complete point cloud down-sampling; finally, the K nearest neighbor algorithm is used to calculate principal curvatures to screen seed points, and region growing is carried out in combination with three-dimensional constraint criteria of normal lines, curvatures and spatial distances to complete defect segmentation and size measurement. The application solves the problems that the existing method is prone to local optimization, error accumulation and loss of small defect features in the process of registration on a blue film weak texture surface, has high registration accuracy, can completely retain defect details, can accurately detect and measure various defects such as pits and scratches, and can meet the needs of industrial high-precision online detection of blue film battery cells.

Description

Technical Field

[0001] This invention relates to a method for detecting defects in the three-dimensional morphology of battery cells, belonging to the field of machine vision and industrial inspection. Background Technology

[0002] As a crucial component of modern high-performance batteries, the quality of the coating on the surface of blue film cells directly determines the core performance and safety of the battery. However, during production and processing, defects such as pits, pinholes, breaks, scratches, or wrinkles inevitably occur on the surface of the blue film. These defects can severely compromise its protective function, leading to internal short circuits, thermal runaway, or accelerated battery performance degradation. Therefore, establishing an efficient and accurate system for detecting and measuring surface defects in blue film cells is of paramount importance.

[0003] In current 3D point cloud detection solutions, point cloud registration is a fundamental step in reconstructing a complete cell membrane model. Traditional iterative nearest-point registration algorithms offer acceptable matching accuracy, but their large iterative computational load and overall high computational complexity make them unsuitable for high-speed online inspection scenarios on production lines. Furthermore, the surface texture of the blue membrane is weak, and the overlap area between multiple viewpoint acquisitions is low, making the algorithm prone to converging to local optima and causing registration misalignment. Registration methods relying on geometric feature points also have significant shortcomings. The overall surface of the blue membrane is smooth, lacking reliably extractable corner points, clear edges, and other distinctive geometric features, often leading to feature extraction failures and the inability to properly stitch together multiple point clouds. Even with the introduction of manual markers to assist registration, this method fragments the overall geometric constraints of the cell, resulting in continuous cumulative errors after multi-frame stitching, significantly reducing the accuracy of 3D reconstruction. In addition, the subsequent point cloud simplification process generally adopts a global uniform downsampling algorithm, which smooths and filters all point cloud data indiscriminately. High-frequency minute details that represent defects, such as pits and fine scratches, are easily lost during data compression, which ultimately greatly reduces the defect detection sensitivity and measurement accuracy, and restricts the reliability of the entire detection system in practical applications. Summary of the Invention

[0004] To address the challenges of point cloud registration due to the lack of significant texture on the surface of blue film battery cells, and the tendency of existing downsampling methods to lose minute defect features, this invention provides a method for detecting three-dimensional morphological defects in battery cells, with the technical solution as follows: Step 1: Acquire images of the battery cells with ArUco markers during the translation process, and complete coarse registration of the battery cell images before and after translation based on the ArUco marker feature points; Step 2: Perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the target point cloud plane, and between the target point cloud and the source point cloud plane, with the goal of minimizing the average distance between the two points, and search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm to complete the fine registration of the point cloud. Step 3: Divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, dynamically adjust the retention ratio of points according to the curvature; for planar grids, only retain the points closest to the centroid of the grid. Step 4: Calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct a three-dimensional constraint criterion including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size.

[0005] Optionally, step 1 includes: Obtain two sets of point clouds containing the same ArUco markers from the images before and after translation. Calculate the 3D centroids of the two point clouds for each set. Subtract the corresponding centroid from the point cloud coordinates to obtain two sets of zero-mean point clouds. Construct the covariance matrix. Singular value decomposition is performed to obtain the optimal rotation matrix and translation vector. The source point cloud is then registered after rigid body transformation by rotation and translation.

[0006] Optionally, step 2 uses a random sampling consistency plane fitting algorithm to fit the corresponding planar regions in the source point cloud and the target point cloud, respectively.

[0007] Optionally, in step 2, the average distance from the point cloud to the plane... and average distance between opposite sides The calculation formula is as follows:

[0008]

[0009] In the formula, This represents a point in a point cloud set. This represents the three-dimensional coordinates of the point. These represent the four parameters of the fitted plane. This represents the total number of points in the point cloud. , , , These represent the source point cloud, source plane, target point cloud, and target plane, respectively.

[0010] Optionally, step 2 searches for the optimal translation amount using a constrained nonlinear minimization optimization algorithm, and the corresponding optimization formula is:

[0011] In the formula, This represents the translation amount of the target plane. This represents the normal vector of the target plane.

[0012] Optionally, the curvature calculation method for points inside the grid in step 3 is as follows:

[0013] In the formula, These are the eigenvalues ​​of the covariance matrix of the neighborhood point set. This represents the curvature at each point.

[0014] Optionally, the three-dimensional constraint criterion is expressed as:

[0015] In the formula, and This represents the seed point and its neighboring points. and The normals between the seed point and its neighboring points are represented. This indicates the angle between the normals of the two lines. For the set included angle threshold, and The principal curvatures of the seed point and its neighboring points Indicates the principal curvature threshold. This represents the Euclidean distance between the seed point and its neighboring points. This represents the distance threshold between two points.

[0016] In a second aspect, the present invention provides a three-dimensional morphological defect detection system for battery cells, the system being configured to implement the three-dimensional morphological defect detection method for battery cells as described in any of the preceding claims, the system comprising: The coarse registration module is configured to acquire cell images with ArUco markers attached during the translation process, and complete the coarse registration of cell images before and after translation based on the ArUco marker feature points; The fine registration module is configured to perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the plane containing the target point cloud, and between the target point cloud and the plane containing the source point cloud, and search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm with the goal of minimizing the average distance between the two points to complete the fine registration of the point cloud. The downsampling processing module is configured to divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, the retention ratio of points is dynamically adjusted according to the curvature; for planar grids, only the points closest to the centroid of the grid are retained. The defect identification module is configured to calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct three-dimensional constraint criteria including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size.

[0017] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the three-dimensional morphology defect detection method for battery cells as described in any of the preceding claims.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the three-dimensional morphology defect detection method for battery cells as described in any of the preceding claims.

[0019] The beneficial effects of this invention are: First, this invention combines ArUco marker features to complete coarse registration of point clouds and relies on planar geometric constraints and constraint nonlinear minimization algorithms to achieve fine registration. This solves the problems of large computational load and easy getting trapped in local optima in traditional iterative nearest point algorithms, as well as the failure of feature extraction due to weak texture of blue film and lack of significant corner edges in conventional geometric feature registration methods, and large cumulative errors in manual marker registration. It effectively improves the registration accuracy of multi-view point clouds. The combination mode of coarse and fine registration also reduces the overall computation time and can be adapted to high-speed online inspection scenarios in industrial production lines.

[0020] Second, the present invention adopts a differentiated downsampling strategy based on curvature-divided grids, and sets different point cloud retention rules for feature grids and planar grids. This solves the problem that the traditional global uniform downsampling algorithm smooths data indiscriminately, causing the loss of high-frequency micro-defect features such as pits and scratches. While greatly simplifying redundant point clouds and improving data processing efficiency, it completely preserves the morphological details of defects on the surface of the battery cell, ensuring the quality of basic data for subsequent defect detection.

[0021] Third, this invention uses the K-nearest neighbor algorithm combined with principal curvature to select defect seed points, and integrates normal smoothness, curvature consistency, and spatial distance to construct three-dimensional constraint criteria for region growth and segmentation. This solves the problems of blind seed point selection and easy missegmentation and merging of regions in traditional defect segmentation algorithms. It achieves accurate segmentation and quantitative measurement of tiny defects in blue film battery cells, effectively reduces the measurement error of various defect parameters, and significantly improves detection sensitivity, accuracy, and stability.

[0022] The overall solution of this invention forms a complete processing link from point cloud registration and data simplification to defect identification and measurement. The whole method is robust and widely adaptable, and can stably identify various blue film surface defects such as protrusions, pits, scratches, and wrinkles. It fully meets the high-precision and high-efficiency quality inspection requirements of power battery cell industrialization and has good engineering application value. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the ArUco marker and its imaging.

[0025] Figure 2 This is a flowchart of a single ArUco marker detection process.

[0026] Figure 3 This is a schematic diagram of the ArUco logo design.

[0027] Figure 4 This is a schematic diagram of the image marker detection results of the present invention.

[0028] Figure 5 This is a flowchart of the raster downsampling algorithm that incorporates curvature features according to the present invention.

[0029] Figure 6 This is a schematic diagram of the point cloud KNN graph model established in this invention.

[0030] Figure 7 This is a physical diagram of the dual-view monocular structured light system and high-precision mobile platform of the present invention.

[0031] Figure 8 This invention relates to an image of a blue film battery cell acquired and a reconstructed point cloud map in a certain pose.

[0032] Figure 9 This is another pose of the blue film battery cell acquisition image and reconstructed point cloud map of the present invention.

[0033] Figure 10 This is a statistical diagram of the coarse registration error of the point cloud of the blue film battery cell in different poses according to the present invention.

[0034] Figure 11 This is a diagram illustrating the convergence process of fine registration optimization in this invention.

[0035] Figure 12 This is a diagram showing the point cloud registration results of blue film battery cells in different poses according to the present invention.

[0036] Figure 13 This is a comparison chart of results from different downsampling methods.

[0037] Figure 14 These are images showing the retention effects of high curvature points in the point cloud after different downsampling methods.

[0038] Figure 15 This is a diagram showing the defect detection and segmentation results of this invention.

[0039] Figure 16 These are illustrations of different types of defects in this invention.

[0040] Figure 17 This is a schematic diagram of a method for measuring the geometric dimensions of defects.

[0041] Figure 18 This is a flowchart of the method for detecting three-dimensional morphological defects in battery cells based on feature geometric constraint point cloud processing, as per the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0043] Example 1: This embodiment provides a method for detecting three-dimensional topographic defects in battery cells based on feature geometric constraint point cloud processing, including: Step 1: Acquire images of the battery cells with ArUco markers during the translation process, and complete coarse registration of the battery cell images before and after translation based on the ArUco marker feature points; Step 2: Perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the target point cloud plane, and between the target point cloud and the source point cloud plane. With minimizing the average distance between the two points as the objective, search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm to complete the fine registration of the point cloud. Step 3: Divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, dynamically adjust the retention ratio of points according to the curvature; for planar grids, only retain the points closest to the centroid of the grid. Step 4: Calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct a three-dimensional constraint criterion including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size.

[0044] This embodiment uses ArUco marker features for coarse registration of point clouds, and combines planar geometric constraints and constraint nonlinearity minimization algorithms to complete fine registration, effectively improving the accuracy and efficiency of multi-view point cloud stitching. Simultaneously, a curvature-driven raster differential downsampling method is introduced to distinguish between feature gratings and planar gratings and employ different retention strategies. This solves the drawback of conventional uniform downsampling, which easily loses minute defect features such as pits and scratches. While reducing the amount of data, it completely preserves surface details. Furthermore, this embodiment relies on the K-nearest neighbor algorithm combined with principal curvature to select seed points, and uses three-dimensional constraint criteria to perform region growth segmentation, achieving accurate extraction and size measurement of defect regions, significantly reducing the probability of missegmentation.

[0045] Example 2: This embodiment provides a method for detecting three-dimensional morphological defects in battery cells based on feature geometric constraint point cloud processing.

[0046] The surface of blue film battery cells has weak texture and no significant geometric features, making it impossible to cover the entire surface with a single structured light scan. Therefore, multi-view point cloud registration and stitching are necessary. Furthermore, minute defects such as pits and scratches are easily lost during downsampling and segmentation, requiring the complete preservation of defect morphology features while simplifying the data. This embodiment achieves high-precision defect detection through ArUco marker coarse registration, planar geometric constraint fine registration, curvature constraint downsampling, and multi-constraint region growth segmentation. The detection method flow of this example can be found in [link to relevant documentation]. Figure 18 This includes the following steps.

[0047] Step 1: Fuse feature points based on external ArUco labels to complete coarse registration.

[0048] The ArUco marker uses a 7×7 module layout, with an outer black border of 1 module width and an internal coding area of ​​5×5 modules, capable of carrying 10 bits of information, corresponding to 1024 ID numbers. The detection process is as follows: Figure 2 As shown, adaptive threshold binarization, Canny edge extraction, closed contour filtering, encoding verification and orientation judgment, and sub-pixel corner optimization are performed to obtain high-precision three-dimensional corner coordinates.

[0049] During actual filming, it will be as follows Figure 3 The marked patterns are affixed to a precision motion platform. The placement of the blue film battery cell is adjusted so that the camera can simultaneously capture a portion of the cell surface and the ArUco markers. After the required pattern for one location area is fully captured, the precision motion platform is used to translate the blue film battery cell within the field of view. During translation, it is necessary to ensure that there is no relative movement between the ArUco markers and the blue film battery cell, and that markers with the same ID number exist in the two images before and after translation for registration. The imaging and marker detection results for two different areas of the blue film battery cell are as follows. Figure 4 As shown.

[0050] After the corner detection of the marked pattern is completed, it is necessary to perform homonymous point matching and retention on the corner point information with the same ID number in two adjacent image regions. Figure 4 The example yielded 12 ArUco tags with the same ID number. Structured light technology was then used to perform 3D reconstruction of the entire image, simultaneously saving the 3D coordinates of the corresponding points. Let the two point cloud sets corresponding to the same points be... and Its expression is as follows:

[0051] In the formula, and Represents the coordinates of a point in the point set.

[0052] Then, the three-dimensional centroids of the two sets of point clouds were calculated respectively. , Eliminate the interference of translation components on registration:

[0053] Subtracting the corresponding centroid from the point cloud coordinates yields a zero-mean point cloud. and Construct the covariance matrix Perform singular value decomposition on the covariance matrix:

[0054]

[0055] The optimal rotation matrix and translation vector obtained are as follows:

[0056] Source Cloud Registration is completed after rigid body transformation involving rotation and translation:

[0057] Coarse registration provides a good initial position, but there may still be minor translation or rotation errors. Fine registration is required to achieve the accuracy needed for detection.

[0058] Step 2: Complete the fine registration based on geometric feature constraints to achieve registration and stitching of point clouds at different locations.

[0059] First, plane fitting and normal vector calculation are performed. The Random Sample Consensus (RANSAC) plane fitting algorithm is used to fit the corresponding planar regions in the source and target point clouds respectively, obtaining their respective plane equations and normal vectors. Then, the initial average distance between the two points is calculated: the average distance from the source point cloud to the plane containing the target point cloud, and the average distance from the target point cloud to the plane containing the source point cloud. The average of these two distances is used as the initial deviation benchmark. (Average distance from point cloud to plane) and average distance between opposite sides The calculation formula is as follows:

[0060]

[0061] In the formula, This represents a point in a point cloud set. This represents the three-dimensional coordinates of the point. These represent the four parameters of the fitted plane. This represents the total number of points in the point cloud. , , , These represent the source point cloud, source plane, target point cloud, and target plane, respectively.

[0062] After calculating the average distance between opposing points, the translation vector is optimized to minimize this average distance. Along the normal vector direction of the target point cloud, a constrained nonlinear minimization optimization algorithm is used to search for the optimal translation amount. The corresponding optimization formula is:

[0063] In the formula, This represents the translation amount of the target plane. This represents the normal vector of the target plane. Iterative optimization of the opposing distance between the two planes ensures symmetry and accuracy during the registration process. Finally, the obtained optimal movement vector is applied to the source point cloud to achieve precise registration of the point cloud.

[0064] Because point cloud registration requires images containing a sufficient number of corresponding feature points, overlapping areas will exist in the resulting point cloud. These overlapping areas contain many redundant points, which reduces computational efficiency and easily introduces noise interference during subsequent processing. Therefore, downsampling and deduplication are necessary while preserving the key geometric information of the point cloud, followed by the detection and segmentation of defects in the point cloud.

[0065] Step 3: Design a grid downsampling algorithm with curvature geometry constraints to remove redundant point clouds while retaining key geometric features of the blue film surface.

[0066] The algorithm flow is as follows Figure 5 As shown, the point cloud space is first divided into grids. Then, the curvature of the point cloud within each grid is calculated. The grid cells are then classified according to a set curvature threshold to determine whether they are feature grids or planar grids. The curvature of points within a grid can be calculated using the following formula.

[0067]

[0068] In the formula, The eigenvalues ​​are the covariance matrix of the neighborhood point set. This represents the curvature at each point. After calculating the curvature at each point, the maximum curvature within the grid is taken.

[0069] In the formula, Represents a point in the grid. This represents the grid formed by dividing a point cloud. This represents the maximum curvature within the grid. If the maximum curvature within a grid is greater than a set curvature threshold, it is classified as a feature grid; otherwise, it is classified as a planar grid.

[0070] In the formula, This indicates the set curvature threshold.

[0071] If a raster is identified as a feature raster, the normals and curvature of its internal point cloud are calculated, and the retention ratio of points is dynamically adjusted based on the curvature. The calculation formula is:

[0072] In the formula, This represents the point cloud retention coefficient, which retains all points in the neighborhood of the feature point to ensure that no details are lost.

[0073] If the grid is determined to be a planar grid, then calculate the grid centroid:

[0074] In the formula, Indicates the grid centroid. , , It represents the component values ​​of the three coordinate directions of each point in the raster, and only the points closest to the centroid of the raster are retained to maximize the refinement of point cloud data.

[0075] By classifying graticules by curvature features and applying different downsampling strategies to point clouds in different graticule types, the details of feature regions can be preserved while reducing the amount of point cloud processing, thus providing data preparation for subsequent point cloud defect detection and segmentation.

[0076] After obtaining the downsampled point cloud, it is necessary to detect and segment the surface defects in the point cloud.

[0077] Step 4: By improving the region growing algorithm and combining it with geometric constraint criteria, the segmentation and dimensional parameter measurement of the micro-defect region are completed.

[0078] In this embodiment, high curvature points are used as the sole source of seed points, and feature localization is achieved through principal curvature calculation.

[0079] First, a K-Nearest Neighbors Graph Model (KNN) is built for each point in the point cloud based on its K nearest neighbors, ensuring that the arrangement of points in the 3D point cloud set remains unchanged. Figure 6 This is a schematic diagram illustrating the creation of a point cloud KNN graph model.

[0080] Then, the first principal curvature of each point in the point cloud is calculated. The calculation shows that the larger the value of the first principal curvature, the higher the degree of local curvature near the point, and the greater the possibility of defects such as bumps and scratches in the point cloud. Then, the curvature histogram is segmented, and points with curvature greater than the threshold are selected as seed points.

[0081] The above method can reduce the number of seed points selected, improve segmentation efficiency, and address the blind seed point selection problem of traditional methods. Based on this method, a three-dimensional constraint criterion of normal smoothness, curvature consistency, and spatial distance is constructed, and accurate clustering of point clouds is achieved through joint judgment of multiple conditions.

[0082] This constraint is first applied through an angle threshold. Restrict the normal vector between the neighboring point and the seed point and The included angle First, the geometric orientation within the cluster is consistent; second, the principal curvature difference is constrained to ensure that the clusters belong to the same type of feature region; finally, the spatial distance range of the neighboring points is limited to maintain the spatial continuity of the cluster, as shown in the following formula.

[0083]

[0084] In the formula, and This represents the seed point and its neighboring points. and The normals between the seed point and its neighboring points are represented. This indicates the angle between the normals of the two lines. For the set included angle threshold, and The principal curvatures of the seed point and its neighboring points Indicates the principal curvature threshold. This represents the Euclidean distance between the seed point and its neighboring points. This represents the distance threshold between two points.

[0085] The aforementioned three-dimensional constraints employ a sequential judgment method. Only neighborhood points that simultaneously satisfy all conditions are added to the current cluster and used as new seed points for further growth. This method avoids over-segmentation of feature regions due to local differences, reduces erroneous merging between background and feature regions, and improves the integrity and independence of the clustering results.

[0086] To verify the technical effectiveness of the method of the present invention, an experiment was conducted. A dual-view monocular structured light system was constructed as follows: Figure 7 As shown, the system includes a camera, projector, and precision motion platform. Results show that the coarse registration RMSE is as low as 0.1254 mm, and the fine registration is optimized to 0.1078 mm; curvature constraint downsampling can preserve high-curvature defect characteristics to the greatest extent; the measurement errors for defect length, width, and depth are all less than 0.03 mm, enabling stable detection of minute defects such as protrusions, pits, and scratches, meeting the accuracy requirements of industrial inspection.

[0087] To verify the effectiveness of this method, a structured light system was used to conduct reconstruction and registration experiments on blue film cells in different poses. Figure 8 and 9 Images and reconstructed point clouds of two different regions of two sets of battery cells with different orientations are shown. The proposed method is used to identify, detect, and pair ArUco markers in the images. Successfully paired corresponding point pairs are shown below. Figure 10 As shown.

[0088] After coarse registration of point clouds using the SVD method, the point pair registration error distribution and final results are as follows: Figure 10 As shown, the registration errors between corresponding point pairs in two different poses are 0.3081 mm and 0.1254 mm, respectively. Applying the transformation matrix to both sets of point clouds yields the coarse registration results as shown below. Figure 11 As shown, the point clouds at two different locations have achieved good initial alignment. The optimization convergence process is as follows: Figure 12 As shown in the figure, after optimization, the overlap of the point cloud was significantly improved, the average distance between opposing points and the RMSE were further reduced, and the fine registration results of the point cloud in the two poses were optimized to 0.1078mm and 0.1242mm, respectively, which met the accuracy requirements of the point cloud model in the subsequent defect detection process.

[0089] To verify the effectiveness of the proposed downsampling method, experiments were conducted using point cloud data registered with blue film cells, and the method was compared with three other classic downsampling methods: random downsampling, uniform downsampling, and voxel grid method. Figure 13 and 14 The results of point cloud downsampling using different methods and the retention effect of high curvature points in the point cloud after downsampling are shown. As can be seen from the figures, this method can preserve the defect features of the blue film battery cell surface to the greatest extent.

[0090] Figure 15 The results of defect detection and segmentation of the downsampled point cloud are shown. Different colored point cloud regions in the figure represent different defect areas. For example... Figure 16 As shown, the proposed method segments three typical defects on the surface of the blue film battery cell: micro-protrusions, pits, and long scratches. Finally, the three-dimensional defect point cloud is projected onto the master plane, and the maximum length and width of the defects are measured on the two-dimensional projection. The depth of the defects is obtained by calculating the distance from the defect point cloud to the reference plane of the normal surface where it resides. A schematic diagram of the defect measurement method is shown below. Figure 17 As shown.

[0091] Example 3: This embodiment provides a three-dimensional morphological defect detection system for battery cells, including: The coarse registration module is configured to acquire cell images with ArUco markers attached during the translation process, and complete the coarse registration of cell images before and after translation based on the ArUco marker feature points; The fine registration module is configured to perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the plane containing the target point cloud, and between the target point cloud and the plane containing the source point cloud, and search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm to complete the fine registration of the point cloud. The downsampling processing module is configured to divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, the retention ratio of points is dynamically adjusted according to the curvature; for planar grids, only the points closest to the centroid of the grid are retained. The defect identification module is configured to calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct three-dimensional constraint criteria including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size.

[0092] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method for detecting three-dimensional morphological defects in battery cells, characterized in that, The method includes: Step 1: Acquire images of the battery cells with ArUco markers during the translation process, and complete coarse registration of the battery cell images before and after translation based on the ArUco marker feature points; Step 2: Perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the target point cloud plane, and between the target point cloud and the source point cloud plane, with the goal of minimizing the average distance between the two points, and search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm to complete the fine registration of the point cloud. Step 3: Divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, dynamically adjust the retention ratio of points according to the curvature; for planar grids, only retain the points closest to the centroid of the grid. Step 4: Calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct a three-dimensional constraint criterion including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size. The average distance from the point cloud to the plane in step 2 and average distance between opposite sides The calculation formula is as follows: In the formula, This represents a point in a point cloud set. This represents the three-dimensional coordinates of the point. These represent the four parameters of the fitted plane. This represents the total number of points in the point cloud. , , , These represent the source point cloud, source plane, target point cloud, and target plane, respectively. Step 2 searches for the optimal translation amount using a constrained nonlinear minimization optimization algorithm. The corresponding optimization formula is: In the formula, This represents the translation amount of the target plane. Represents the normal vector of the target plane; The method for calculating the curvature of points inside the grid in step 3 is as follows: In the formula, These are the eigenvalues ​​of the covariance matrix of the neighborhood point set. Indicates the curvature at each point; The three-dimensional constraint criterion is expressed as follows: In the formula, and Represents the seed point and its neighboring points. and The normals between the seed point and its neighboring points are represented. This indicates the angle between the normals of the two lines. For the set included angle threshold, and The principal curvatures of the seed point and its neighboring points Indicates the principal curvature threshold. This represents the Euclidean distance between the seed point and its neighboring points. This represents the distance threshold between two points.

2. The method for detecting three-dimensional morphological defects in battery cells according to claim 1, characterized in that, Step 1 includes: Obtain two sets of point clouds containing the same ArUco markers from the images before and after translation. Calculate the 3D centroids of the two point clouds for each set. Subtract the corresponding centroid from the point cloud coordinates to obtain two sets of zero-mean point clouds. Construct the covariance matrix. Singular value decomposition is performed to obtain the optimal rotation matrix and translation vector. The source point cloud is then registered after rigid body transformation by rotation and translation.

3. The method for detecting three-dimensional morphological defects in battery cells according to claim 1, characterized in that, Step 2 uses a random sampling consistency plane fitting algorithm to fit the corresponding planar regions in the source point cloud and the target point cloud, respectively.

4. A three-dimensional morphology defect detection system for battery cells, characterized in that, The system is configured to implement the cell three-dimensional morphology defect detection method as described in any one of claims 1-3, the system comprising: The coarse registration module is configured to acquire cell images with ArUco markers attached during the translation process, and complete the coarse registration of cell images before and after translation based on the ArUco marker feature points; The fine registration module is configured to perform plane fitting on the coarsely registered point cloud, calculate the average distance between the source point cloud and the plane containing the target point cloud, and between the target point cloud and the plane containing the source point cloud, and search for the optimal translation amount along the normal vector direction using a constrained nonlinear minimization optimization algorithm with the goal of minimizing the average distance between the two points to complete the fine registration of the point cloud. The downsampling processing module is configured to divide the finely registered point cloud into spatial grids, calculate the curvature of the point cloud within each grid, and divide the grid into feature grids and planar grids based on the curvature threshold; for feature grids, the retention ratio of points is dynamically adjusted according to the curvature; for planar grids, only the points closest to the centroid of the grid are retained. The defect identification module is configured to calculate the first principal curvature of the point cloud based on the K-nearest neighbor algorithm, select points with curvature greater than a preset threshold as seed points, construct three-dimensional constraint criteria including normal smoothness, curvature consistency and spatial distance, and perform clustering and segmentation of seed points through a region growing algorithm to extract defect regions and measure their size.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the cell three-dimensional morphology defect detection method as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the cell three-dimensional morphology defect detection method as described in any one of claims 1 to 3.

Images