Cylindrical battery pole piece coating uniformity detection method based on machine vision

Abstract

The application provides a cylindrical battery pole piece coating uniformity detection method based on machine vision, relates to the technical field of machine vision, and comprises the following steps: acquiring multi-angle images and carrying out cylindrical surface development; performing polar coordinate gradient continuity analysis on the annular development image, extracting the continuity feature at the joint, and adaptively adjusting the light-emitting parameters of an LED array to obtain an optimized image; extracting coating edge profile, surface texture directionality and density distribution features in the polar coordinate domain and the Cartesian coordinate domain, and establishing a coating uniformity evaluation standard; constructing a hierarchical feature fusion network comprising a polar coordinate attention module and a spatial attention module, adaptively weighting and fusing annular features and axial features, comparing the fusion result with a theoretical coating uniformity distribution calculated based on coating process parameters, and outputting a detection result. The application improves the accuracy and interpretability of the cylindrical battery pole piece coating uniformity detection, and provides a theoretical basis for process optimization.

Description

Technical Field

[0001] This invention relates to machine vision technology, and more particularly to a machine vision-based method for detecting the coating uniformity of cylindrical battery electrode sheets. Background Technology

[0002] The uniformity of electrode coating in cylindrical batteries is a key factor affecting battery performance and safety. Existing detection methods have three main limitations: traditional machine vision systems struggle to handle curvature variations and uneven illumination on cylindrical surfaces, especially at seams where image distortion is common; conventional feature extraction algorithms, designed in Cartesian coordinates, cannot effectively capture the periodic features and radial variations of cylindrical surfaces, resulting in incomplete feature representation; and there is a lack of evaluation mechanisms that combine multi-dimensional features with theoretical coating models, leaving detection results without theoretical support and making it difficult to trace back to specific process parameters.

[0003] These technical limitations result in low accuracy of existing detection methods, particularly when dealing with complex coating defects. Furthermore, the detection results lack correlation with actual coating process parameters, hindering process optimization and quality control. Additionally, fixed-parameter image acquisition schemes struggle to adapt to varying environmental conditions and changes in battery electrode surface characteristics, leading to poor stability of the detection results. Therefore, there is an urgent need for a coating uniformity detection method that can adapt to the characteristics of cylindrical surfaces, incorporates theoretical models, and possesses adaptive capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a machine vision-based method for detecting the coating uniformity of cylindrical battery electrode sheets, which can solve the problems in existing technologies, including:

[0005] LED arrays are arranged along the circumferential and axial directions of the cylindrical battery electrode to acquire multi-angle images and unfold the cylindrical surface to obtain circumferential unfolded images and axial unfolded images.

[0006] Polar coordinate gradient continuity analysis is performed on the circumferential unfolded image to extract the continuity features at the seams. Based on the continuity features at the seams, the LED array luminous parameters are adaptively adjusted and the multi-angle images are reacquired to obtain an optimized circumferential unfolded image.

[0007] Feature extraction is performed on the optimized circumferential unfolded image and the axial unfolded image in polar coordinate domain and Cartesian coordinate domain respectively to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features. Based on the coating edge contour features, a coating edge integrity index is calculated; based on the coating surface texture directionality features, a coating direction consistency index is calculated; and based on the coating density distribution features, a coating density uniformity index is calculated to establish a coating uniformity evaluation standard.

[0008] A hierarchical feature fusion network is constructed, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion. The weight coefficients of circumferential and axial features are calculated according to the coating uniformity evaluation criteria. The circumferential and axial features are adaptively weighted and fused. The fusion result is compared with the theoretical coating uniformity distribution, and the coating uniformity detection result is output. The theoretical coating uniformity distribution is calculated based on coating process parameters.

[0009] In one alternative implementation,

[0010] The steps of performing polar coordinate gradient continuity analysis on the circumferential unfolded image, extracting continuity features at the seams, adaptively adjusting the LED array luminous parameters based on the continuity features at the seams, and re-acquiring the multi-angle images include:

[0011] An adaptive polar coordinate mapping relationship is established for the circumferential unfolded image, and the grid density is dynamically adjusted according to the curvature change of the cylindrical surface to obtain a polar coordinate domain image; the theoretical reflection intensity distribution map of the polar coordinate domain image is calculated.

[0012] The polar coordinate domain image is compared with the theoretical reflection intensity distribution map. The reflection intensity difference map is divided into multiple sub-regions using a superpixel segmentation algorithm. The radial gradient and angular gradient of the sub-regions are calculated. A hierarchical continuity evaluation index is constructed by combining local binary pattern features to obtain the continuity score of each sub-region.

[0013] A target optimization function is constructed based on the continuity score. The target optimization function includes a weighted sum of seam continuity error, reflection uniformity error, and edge sharpness error. Based on the target optimization function, a particle swarm optimization algorithm is used to iteratively optimize the luminous parameters of the LED array. The particle swarm optimization algorithm is combined with a simulated annealing strategy to dynamically adjust the particle velocity and obtain the initial luminous parameters of the LED array.

[0014] A self-calibrating marker array is preset in the circumferential unfolded image. The reflection characteristics of the marker array are collected to obtain the environmental change feature vector. A polynomial mapping relationship between the environmental change feature vector and the LED compensation parameters is established, and the real-time LED compensation parameters are calculated. The compensation parameters are applied to the initial light emission parameters to obtain the optimized LED array light emission parameters. The image is re-acquired using the optimized LED array light emission parameters.

[0015] In one alternative implementation,

[0016] The steps involved in dividing the reflection intensity difference map into multiple sub-regions using a superpixel segmentation algorithm, calculating the radial and angular gradients of each sub-region, and constructing a hierarchical continuity evaluation index by combining local binary pattern features include:

[0017] The reflection intensity difference map is preprocessed. A polar coordinate grid is established in the preprocessed reflection intensity difference map based on the surface curvature of the cylindrical battery electrode. The weighted response values ​​of radial gradient and angular gradient are calculated. The seed point position for superpixel segmentation is determined based on the weighted response values. A distance metric function containing spatial distance and color distance terms is constructed with the seed point as the center. Based on the distance metric function, region growing is performed on the preprocessed reflection intensity difference map to obtain multiple sub-regions.

[0018] Multi-scale radial gradient features are obtained by convolution operation on the sub-region using the multi-scale Sobel operator. The angular gradient features of the sub-region are calculated in polar coordinates. The dynamic fusion weight of the radial gradient features and angular gradient features is determined according to the local curvature of the cylindrical battery electrode to obtain the gradient features of the sub-region.

[0019] Local binary pattern features are extracted from the sub-region using multiple circular templates with different sampling radii and sampling point numbers. These local binary pattern features are then mapped to rotation-invariant equivalent patterns to obtain the texture features of the sub-region.

[0020] The area, perimeter, roundness, and eccentricity of the sub-regions are extracted as morphological features;

[0021] The gradient features, texture features, and morphological features are constructed into a hierarchical feature system, and the continuity score of the sub-region is calculated based on the hierarchical feature system.

[0022] In one alternative implementation,

[0023] The steps of extracting features from the optimized circumferential unfolded image and the axial unfolded image in the polar coordinate domain and the Cartesian coordinate domain, respectively, to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features include:

[0024] An adaptive Gabor filter is constructed in the polar coordinate domain. The radial and angular standard deviations of the adaptive Gabor filter are adaptively adjusted according to the curvature of the cylindrical battery electrode, and the optimized circumferential unfolded image and the axial unfolded image are filtered and enhanced respectively.

[0025] Based on the enhanced optimized circumferential unfolded image and axial unfolded image, a curvature scale space is constructed. The curvature descriptor of the coating edge is calculated under different Gaussian smoothing scales. The curvature descriptor contains the first and second derivatives of the contour point coordinates, and the contour features of the coating edge are obtained.

[0026] The gradients of the enhanced optimized circumferential unfolded image and axial unfolded image are calculated in the Cartesian coordinate domain to construct a structural tensor matrix. The structural tensor matrix is ​​then decomposed into eigenvalues ​​to obtain principal direction eigenvalues. Based on the principal direction eigenvalues, the directional features of the coating surface texture are calculated.

[0027] An adaptive kernel density function is used to estimate the density of the enhanced optimized circumferential and axial unfolded images. The local bandwidth of the adaptive kernel density function is dynamically adjusted according to the gray-level distribution of the coating area to construct density feature vectors at multiple scales and obtain the coating density distribution features.

[0028] In one alternative implementation,

[0029] The steps for establishing a coating uniformity evaluation standard include: calculating a coating edge integrity index based on the coating edge contour features, calculating a coating direction consistency index based on the coating surface texture directionality features, and calculating a coating density uniformity index based on the coating density distribution features.

[0030] The coating edge contour features are segmented and fitted, and the curvature continuity score and contour integrity score are calculated to generate a coating edge integrity index.

[0031] Establish a directional feature evaluation criterion, calculate the main direction consistency coefficient of the local area and the global direction distribution entropy value based on the directional features of the coating surface texture, and generate a coating direction consistency index;

[0032] Design a density distribution evaluation function, calculate the regional density mean deviation and density gradient change rate based on the coating density distribution characteristics, and generate a coating density uniformity index.

[0033] The coating edge integrity index, coating direction consistency index, and coating density uniformity index are normalized using a fuzzy comprehensive evaluation method to establish a coating uniformity evaluation standard.

[0034] In one alternative implementation,

[0035] The steps of constructing a hierarchical feature fusion network, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion, calculating the weight coefficients of circumferential and axial features according to the coating uniformity evaluation criteria, adaptively weighting and fusing the circumferential and axial features, comparing the fusion result with the theoretical coating uniformity distribution, and outputting the coating uniformity detection result include:

[0036] A polar coordinate feature map is constructed from the features extracted in the polar coordinate domain of the optimized circumferential unfolded image, and a Cartesian coordinate feature map is constructed from the features extracted in the Cartesian coordinate domain of the axial unfolded image.

[0037] The polar coordinate attention module constructs a periodic projection matrix based on the angle information in the polar coordinate feature map. The periodic projection matrix maps the polar coordinate feature map to a query vector, a key vector, and a value vector. The product of the query vector and the key vector is calculated, divided by a preset scaling factor, and then passed through a softmax function to obtain the attention weight. The product of the attention weight and the value vector is used as the polar coordinate attention feature.

[0038] The spatial attention module performs convolution, pooling, and multilayer perceptron operations sequentially on the input Cartesian coordinate feature map to obtain spatial attention features.

[0039] Based on the coating uniformity evaluation standard, the circumferential feature weight coefficient and the axial feature weight coefficient are calculated. The circumferential feature weight coefficient and the axial feature weight coefficient are multiplied by the polar coordinate attention feature and the spatial attention feature respectively and then superimposed to obtain the fused feature.

[0040] A theoretical coating uniformity distribution model was established based on coating process parameters;

[0041] Calculate the absolute sum of the differences between the fusion feature and the theoretical coating uniformity distribution model, divide the absolute sum of the differences by the integral value of the theoretical coating uniformity distribution model to obtain the normalized difference degree, and output the coating uniformity detection result based on the normalized difference degree.

[0042] In one alternative implementation,

[0043] The steps for establishing a theoretical coating uniformity distribution model based on coating process parameters include:

[0044] The coating process parameters include coating slurry viscosity, coating speed, and coating pressure;

[0045] A linear correction relationship between the viscosity of the coating slurry and temperature is established. For a given reference temperature and reference viscosity, a viscosity change coefficient corresponding to a unit temperature change is established to obtain the temperature correction value. Based on the average shear rate in the actual coating process, a viscosity shear thinning correction coefficient table is established, and the shear correction coefficient is obtained by looking up the table. The viscosity correction value is obtained by multiplying the temperature correction value by the shear correction coefficient.

[0046] A piecewise linear function is used to describe the relationship between coating pressure and coating thickness. The coating area is divided into an inlet section, an intermediate section and an outlet section. The thickness correction coefficient of each section is calculated based on the pre-calibrated pressure-thickness correspondence.

[0047] The curvature coefficient is calculated based on the circumferential angle and radial position of the cylindrical battery electrode. The curvature coefficient is then combined with the cosine function of the circumferential angle and the linear function of the radial position to obtain a geometric compensation value.

[0048] The viscosity correction value, thickness correction coefficient, coating speed, and geometric compensation value are combined by weighted averaging to generate a theoretical coating uniformity distribution model.

[0049] This invention achieves distortion-free imaging of cylindrical surfaces from multiple angles through a bidirectional LED array arrangement in both circumferential and axial directions and adaptive parameter adjustment technology, effectively solving the image distortion problem at curved surface seams. The combination of polar coordinate gradient continuity analysis and a self-calibrating marker array significantly improves image quality and environmental adaptability. Simultaneously, a dual-domain feature extraction method in both polar and Cartesian coordinates accurately captures coating edges, texture direction, and density distribution features, resulting in a more comprehensive feature representation.

[0050] This invention innovatively constructs a hierarchical feature fusion network, using polar coordinate attention modules and spatial attention modules to process circumferential and axial features respectively, achieving effective fusion of periodic and spatial features. Comparing the detection results with a theoretical coating uniformity distribution model established based on coating process parameters not only improves detection accuracy but also establishes a correlation between visual detection results and process parameters, providing a theoretical basis for process optimization, while simultaneously achieving interpretability and traceability of the detection results. Attached Figure Description

[0051] Figure 1 This is a schematic flowchart of a machine vision-based method for detecting the coating uniformity of cylindrical battery electrode sheets according to an embodiment of the present invention.

[0052] Figure 2 This is a flowchart of superpixel segmentation and hierarchical continuity evaluation. Detailed Implementation

[0053] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a schematic flowchart of the machine vision-based method for detecting the coating uniformity of cylindrical battery electrodes according to the present invention. Figure 1 As shown, the method includes:

[0055] LED arrays are arranged along the circumferential and axial directions of the cylindrical battery electrode to acquire multi-angle images and unfold the cylindrical surface to obtain circumferential unfolded images and axial unfolded images.

[0056] Polar coordinate gradient continuity analysis is performed on the circumferential unfolded image to extract the continuity features at the seams. Based on the continuity features at the seams, the LED array luminous parameters are adaptively adjusted and the multi-angle images are reacquired to obtain an optimized circumferential unfolded image.

[0057] Feature extraction is performed on the optimized circumferential unfolded image and the axial unfolded image in polar coordinate domain and Cartesian coordinate domain respectively to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features. Based on the coating edge contour features, a coating edge integrity index is calculated; based on the coating surface texture directionality features, a coating direction consistency index is calculated; and based on the coating density distribution features, a coating density uniformity index is calculated to establish a coating uniformity evaluation standard.

[0058] A hierarchical feature fusion network is constructed, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion. The weight coefficients of circumferential and axial features are calculated according to the coating uniformity evaluation criteria. The circumferential and axial features are adaptively weighted and fused. The fusion result is compared with the theoretical coating uniformity distribution, and the coating uniformity detection result is output. The theoretical coating uniformity distribution is calculated based on coating process parameters.

[0059] In one alternative implementation,

[0060] The steps of performing polar coordinate gradient continuity analysis on the circumferential unfolded image, extracting continuity features at the seams, adaptively adjusting the LED array luminous parameters based on the continuity features at the seams, and re-acquiring the multi-angle images include:

[0061] An adaptive polar coordinate mapping relationship is established for the circumferential unfolded image, and the grid density is dynamically adjusted according to the curvature change of the cylindrical surface to obtain a polar coordinate domain image; the theoretical reflection intensity distribution map of the polar coordinate domain image is calculated.

[0062] The polar coordinate domain image is compared with the theoretical reflection intensity distribution map. The reflection intensity difference map is divided into multiple sub-regions using a superpixel segmentation algorithm. The radial gradient and angular gradient of the sub-regions are calculated. A hierarchical continuity evaluation index is constructed by combining local binary pattern features to obtain the continuity score of each sub-region.

[0063] A target optimization function is constructed based on the continuity score. The target optimization function includes a weighted sum of seam continuity error, reflection uniformity error, and edge sharpness error. Based on the target optimization function, a particle swarm optimization algorithm is used to iteratively optimize the luminous parameters of the LED array. The particle swarm optimization algorithm is combined with a simulated annealing strategy to dynamically adjust the particle velocity and obtain the initial luminous parameters of the LED array.

[0064] A self-calibrating marker array is preset in the circumferential unfolded image. The reflection characteristics of the marker array are collected to obtain the environmental change feature vector. A polynomial mapping relationship between the environmental change feature vector and the LED compensation parameters is established. The real-time LED compensation parameters are calculated based on the polynomial mapping relationship. The compensation parameters are applied to the initial light emission parameters to obtain the optimized LED array light emission parameters. The image is re-acquired using the optimized LED array light emission parameters. This process is repeated until the target optimization function value is less than a preset threshold.

[0065] In this embodiment, the process of establishing an adaptive polar coordinate mapping relationship for the circumferentially unfolded image involves curvature-adaptive mesh construction. The circumferentially unfolded image is represented in a Cartesian coordinate system. To establish an analysis model suitable for the cylindrical surface, it needs to be mapped to the polar coordinate domain. Considering the curvature variation of the cylindrical battery electrode surface, the mesh density is increased in areas with larger curvature and appropriately reduced in areas with smaller curvature. Specifically, the width of the circumferentially unfolded image is set to W pixels and the height to H pixels, and a temporary Cartesian coordinate system is established with the image center as the origin. For the case of a cylindrical battery with a diameter of 18mm, sampling is performed at 1-degree angle intervals, with a variable step size in the radial direction, from the inside out: 0.1mm, 0.15mm, 0.2mm, and 0.25mm, respectively, to ensure denser sampling points in areas with large curvature variations on the battery electrode surface (such as the area near the battery's central axis). When each pixel (x, y) is converted to polar coordinates (r, θ), r represents the distance to the origin, and θ represents the angle with the horizontal axis. The polar coordinate domain image is obtained through point-by-point transformation.

[0066] After acquiring the polar coordinate image, the light intensity distribution of the LED array and the reflectivity of the coating material were calculated. The light intensity distribution of the LED array was described using the Lambertian model, with the intensity varying with the cosine of the angle centered on the LED light source. For an 8×8 LED array, with a spacing of 5mm between each LED, a power of 0.5W, and an initial emission angle of 45 degrees, the spatial light intensity distribution was obtained. The reflectivity of the coating material was obtained by measuring the reflectivity at different incident angles. A typical lithium battery electrode coating has a reflectivity of 0.25 at perpendicular incidence, which increases with increasing incident angle, reaching 0.38 at an incident angle of 60 degrees. Combining the LED light intensity distribution and the material reflectivity, the received and reflected light intensities at each sampling point were calculated point by point to generate a theoretical reflection intensity distribution map.

[0067] The polar coordinate image is compared pixel-by-pixel with the theoretical reflectance intensity distribution map, and the difference is calculated to generate a reflectance intensity difference map. This difference map is then divided into semantically consistent sub-regions using a superpixel segmentation algorithm. The SLIC algorithm is employed; for a 1024×768 image, the initial number of superpixels is set to 100, and the compactness factor is 10. The superpixel segmentation result is a set of sub-regions of similar size with well-preserved boundaries. For each sub-region, radial and angular gradients are calculated. The radial gradient is represented by the average gray-level difference between adjacent pixels along the radial direction, and the angular gradient is represented by the average gray-level difference between adjacent pixels along the angular direction. In practical applications, for electrode images of 18mm diameter cylindrical batteries, the ideal radial gradient value should be less than 5 gray levels / mm, and the angular gradient value should be less than 3 gray levels / degree.

[0068] In the local binary pattern feature extraction process, circular sampling templates with radii of 1, 2, and 3 pixels are selected, with a sampling point count of 8. The LBP value is calculated for each pixel within each sub-region, and an LBP histogram is generated. The histogram is normalized to obtain the texture features of the sub-region. A hierarchical continuity evaluation index is constructed by combining radial gradient, angular gradient, and LBP features. This index includes three levels: low-level gradient continuity, mid-level texture consistency, and high-level region similarity. The weight for low-level gradient continuity is 0.5, for mid-level texture consistency it is 0.3, and for high-level region similarity it is 0.2. Through weighted fusion, a continuity score for each sub-region is obtained, ranging from 0 to 100, with higher values ​​indicating better continuity.

[0069] The objective optimization function is constructed based on continuity scores. The seam continuity error is the squared difference between the continuity score of the sub-region at the seam and the ideal score (set to 90). The reflection uniformity error is the standard deviation of the continuity scores of all sub-regions. The edge sharpness error is the squared difference between the gradient value of the edge region and the expected gradient value (set to 8 gray levels / mm). The weighting coefficients for the three errors are 0.4, 0.35, and 0.25, respectively, forming a weighted sum objective optimization function.

[0070] The particle swarm optimization algorithm was used to iteratively optimize the luminous parameters of the LED array. Thirty particles were initialized, each representing a set of LED parameters (including brightness, angle, etc.). The particle position ranged from 0-255 for brightness and 0-90 degrees for angle, with initial velocities ranging from ±20 for brightness and ±5 for angle. The optimal individual and global positions were initialized based on the results of the first iteration. Each iteration updated the particle velocity and position, incorporating inertia weights, cognitive coefficients, and social coefficients of 0.7, 1.5, and 1.5, respectively. A simulated annealing strategy was employed, setting the initial temperature to 100°C and the cooling coefficient to 0.95. The temperature was gradually decreased with each iteration, resulting in a smaller change in particle velocity, thus obtaining the initial luminous parameters of the LED array.

[0071] To adapt to environmental changes, a self-calibrating marker array is pre-set in the circumferentially unfolded image. The marker array consists of five points evenly distributed along the circumference of the battery electrode, each with a diameter of 1 mm, made of gray cardstock with a standard reflectivity of 0.18. The reflectivity of the marker array is collected in real time, and the average gray value and standard deviation of the five markers are calculated to form an environmental change feature vector. Through multiple tests, a third-order polynomial mapping relationship between the environmental feature vector and the LED compensation parameters is established. For example, the brightness compensation value is negatively correlated with the average gray value and positively correlated with the standard deviation of the gray value. Specific mapping coefficients are obtained through least-squares fitting; for example, when the ambient light increases by 10%, the LED brightness needs to be reduced by 5-8%; when the ambient temperature increases by 5°C, the LED angle needs to be adjusted by 2-3 degrees.

[0072] The real-time LED compensation parameters are applied to the initial luminous parameters of the LED array. During the compensation process, the brightness parameter is compensated using multiplication, and the angle parameter is compensated using addition. The circumferential unfolded image is reacquired using the optimized parameters, and the target optimization function value is recalculated. If the target function value is less than the preset threshold (set to 50), the optimization process is complete; otherwise, parameter optimization and image acquisition continue.

[0073] This invention accurately captures the surface characteristics of cylindrical battery electrodes through adaptive polar coordinate mapping and superpixel segmentation technology; it effectively solves the problem of uneven illumination in curved surface imaging by combining particle swarm optimization algorithm and simulated annealing strategy for LED parameter optimization; and it improves adaptability in complex environments through self-calibrated marker array design and environmental compensation mechanism.

[0074] In one alternative implementation,

[0075] The steps involved in dividing the reflection intensity difference map into multiple sub-regions using a superpixel segmentation algorithm, calculating the radial and angular gradients of each sub-region, and constructing a hierarchical continuity evaluation index by combining local binary pattern features include:

[0076] The reflection intensity difference map is preprocessed. A polar coordinate grid is established in the preprocessed reflection intensity difference map based on the surface curvature of the cylindrical battery electrode. The weighted response values ​​of radial gradient and angular gradient are calculated. The seed point position for superpixel segmentation is determined based on the weighted response values. A distance metric function containing spatial distance and color distance terms is constructed with the seed point as the center. Based on the distance metric function, region growing is performed on the preprocessed reflection intensity difference map to obtain multiple sub-regions.

[0077] Multi-scale radial gradient features are obtained by convolution operation on the sub-region using the multi-scale Sobel operator. The angular gradient features of the sub-region are calculated in polar coordinates. The dynamic fusion weight of the radial gradient features and angular gradient features is determined according to the local curvature of the cylindrical battery electrode to obtain the gradient features of the sub-region.

[0078] Local binary pattern features are extracted from the sub-region using multiple circular templates with different sampling radii and sampling point numbers. These local binary pattern features are then mapped to rotation-invariant equivalent patterns to obtain the texture features of the sub-region.

[0079] The area, perimeter, roundness, and eccentricity of the sub-regions are extracted as morphological features;

[0080] The gradient features, texture features, and morphological features are constructed into a hierarchical feature system, and the continuity score of the sub-region is calculated based on the hierarchical feature system.

[0081] Combination Figure 2 Superpixel segmentation and hierarchical continuity evaluation flowchart: In this embodiment, for the acquired reflection intensity difference map, high-frequency noise is first removed by Gaussian filtering, with the filter kernel size set to 5×5 pixels and the standard deviation set to 1.2. Then, contrast stretching is performed, linearly mapping the grayscale value range from the original [25, 180] to [0, 255] to enhance image contrast. Gamma correction is applied to highlight areas (grayscale values ​​greater than 200), with a gamma value set to 0.8 to avoid overexposure. Adaptive histogram equalization is applied to shadow areas (grayscale values ​​less than 50), with a block size of 16×16 pixels and a contrast limit parameter of 2.5 to improve details in dark areas. When establishing a polar coordinate grid for the preprocessed image, the surface curvature of the cylindrical battery electrode is considered. With the central axis of the cylinder as the origin, the radial resolution is set to 0.1mm in the paraxial region (0-5mm), 0.15mm in the middle region (5-8mm), and 0.2mm in the outer region (above 8mm); the angular resolution is uniformly set to 1 degree.

[0082] The weighted response values ​​of the radial and angular gradients are calculated on a polar coordinate grid. The radial gradient is calculated using the gray-level difference between adjacent grid points along the radial direction, and the angular gradient is calculated using the gray-level difference between adjacent grid points along the angular direction. In the radial direction, the gradient is calculated using forward difference. For a pixel at position (r, θ), its radial gradient is the difference between the gray-level value of that point and the gray-level value of (r + Δr, θ) divided by Δr. In the angular direction, the difference between the gray-level values ​​of (r, θ) and (r, θ + Δθ) is calculated divided by r·Δθ. Δr represents the distance increment between radially adjacent grid points, and Δθ represents the angular increment between angularly adjacent grid points. Based on the characteristics of the cylindrical surface, the radial gradient weight coefficient is set to 0.6, and the angular gradient weight coefficient is set to 0.4. The weighted response values ​​are then calculated.

[0083] Seed point locations for superpixel segmentation are determined based on weighted response values. An initial number of seed points is set to 100, and these initial seed points are evenly distributed across the reflectance intensity difference map. The seed point locations are adjusted using a gradient suppression strategy: if a seed point is located in a high-gradient region (weighted response value greater than a threshold of 25), it is moved towards surrounding low-gradient regions. The moving distance is proportional to the gradient value, with a maximum moving distance limited to 1 / 3 of the original seed point spacing. After adjustment, seed points tend to be distributed more evenly across the image, improving the stability of subsequent segmentation. For an image with a resolution of 1024×768, the initial spacing between seed points is approximately 80 pixels, the adjusted average spacing is approximately 75 pixels, and the minimum spacing is no less than 60 pixels.

[0084] A distance metric function is constructed centered on the adjusted seed point. This function includes spatial and color distance terms. Spatial distance is the Euclidean distance from the pixel to the seed point, and color distance is the absolute value of the difference in grayscale values ​​between the pixel and the seed point. The weight coefficient m for the spatial distance term is set to 20 to control the compactness of the superpixel shape; the weight for the color distance term is 1. The value of m can be dynamically adjusted for different regions: in seam regions, m is reduced to 15 to increase the influence of color similarity; in uniform regions, m is increased to 25 to make the superpixel shape more regular. Based on this distance metric function, a region growing algorithm is executed starting from the seed point. In each iteration, each superpixel expands outward by one pixel, incorporating the nearest neighboring pixel into the current superpixel. The iteration process continues until all pixels in the image have been assigned to a superpixel.

[0085] Boundary optimization is performed on the formed sub-regions to eliminate small regions while maintaining smooth boundaries. Superpixels with an area less than 1% of the total number of pixels are removed, and their pixels are redistributed to neighboring superpixels. Boundary smoothing employs morphological operations, using circular structuring elements with a radius of 2 for opening operations to remove small protrusions. The final number of superpixels is typically between 80 and 90, with an average area of ​​approximately 8000-9000 pixels. Feature extraction is performed on each sub-region. First, a multi-scale Sobel operator is used to calculate radial gradient features. Three scales of Sobel operators—3×3, 5×5, and 7×7—are designed to extract radial edge information at different scales. The 3×3 operator captures fine edges, the 5×5 operator extracts medium-scale features, and the 7×7 operator captures large-scale changes. These three operators are applied pixel-by-pixel to the sub-regions to calculate gradient magnitudes, and gradient magnitude histograms are plotted within each sub-region. The histograms are divided into 10 bins and normalized to form a 30-dimensional feature vector. In images of 18650 batteries, the proportion of low-gradient bins (1-3) in the multi-scale radial gradient feature vector of normal coating areas typically exceeds 65%, while the proportion of high-gradient bins (7-10) in defect areas can reach over 25%. When calculating the angular gradient features of sub-regions in polar coordinates, the central difference method is used. For each pixel (r, θ) within a sub-region, half the difference between the gray values ​​of (r, θ + Δθ) and (r, θ - Δθ) is divided by r·Δθ to obtain the angular gradient. The angular gradient values ​​are also statistically analyzed into a 10-bin histogram, forming a 10-dimensional feature vector.

[0086] The local curvature calculation of cylindrical battery electrodes is based on the distance from the electrode surface to the central axis. In the near-axis region (large curvature), the angular gradient weight increases to 0.6, while the radial gradient weight decreases to 0.4. In the far-axis region (small curvature), the angular gradient weight decreases to 0.3, while the radial gradient weight increases to 0.7. Dynamic weights are applied to weighted fuse the multi-scale radial and angular gradient features to obtain the comprehensive gradient features of the sub-region.

[0087] When extracting local binary pattern features from sub-regions, multiple circular sampling templates are designed. Circular templates with radii of 1, 2, and 3 pixels are used, with sampling points of 8, 12, and 16 respectively. For each pixel within the sub-region, its local binary pattern value is extracted. Specifically, the grayscale value of each sampling point within the circular neighborhood is compared with the grayscale value of the center pixel; values ​​greater than the center pixel are set to 1, and values ​​less than or equal to the center pixel are set to 0, forming a binary code. For example, using a template with a radius of 1 pixel and 8 sampling points might generate an 8-bit binary code like 00011100. To improve rotation invariance, all binary codes are cyclically shifted to the minimum value as the equivalent pattern. For example, the equivalent pattern of 00011100 and 00111000 is both 00011100.

[0088] Equivalent patterns of all pixels within a sub-region are statistically analyzed to form a histogram. There are 59 equivalent pattern types with a radius of 1 pixel and 8 sampling points; 135 equivalent pattern types with a radius of 2 pixels and 12 sampling points; and 243 equivalent pattern types with a radius of 3 pixels and 16 sampling points. These three histograms are merged to form a 437-dimensional texture feature vector. In practical applications, dimensionality reduction can be performed to select the 100 most discriminative features. For normal coating areas, texture features exhibit strong directionality and good regularity, with a concentrated distribution of equivalent patterns; while in defective areas, the distribution of equivalent patterns is more dispersed, and the entropy value is higher.

[0089] The morphological features of the sub-regions are extracted, including area, perimeter, roundness, and eccentricity. Area is calculated as the number of pixels within the sub-region; perimeter is the number of pixels at the sub-region boundary; roundness is defined as 4π times the area divided by the square of the perimeter, with a perfect circle having a roundness of 1; eccentricity is calculated as the ratio of the principal axis length to the secondary axis length of the sub-region, representing the extensibility of the region's shape. These four morphological features form a 4-dimensional feature vector. For standard coated regions, roundness is typically between 0.7 and 0.9, and eccentricity is between 1.1 and 1.5; while defective regions often have roundness below 0.6 and eccentricity potentially above 2.0.

[0090] A hierarchical feature system is constructed using gradient features, texture features, and morphological features. The lower layer consists of 40-dimensional gradient features, the middle layer of 100-dimensional texture features, and the upper layer of 4-dimensional morphological features. In this hierarchical feature system, lower-level features reflect local changes, middle-level features capture structural patterns, and upper-level features describe the overall morphology. When calculating the continuity score of a sub-region based on this feature system, scores for each of the three layers are calculated separately and assigned different weights. The weight for lower-level gradient features is 0.5, for middle-level texture features it is 0.3, and for upper-level morphological features it is 0.2. For lower-level gradient features, the score is calculated based on their similarity to a preset ideal gradient distribution; for middle-level texture features, the score is evaluated based on their cross-correlation coefficient with a standard template; and for upper-level morphological features, the score is evaluated based on their deviation from the ideal morphological parameters. The weighted fusion of the three-layer scores yields the final continuity score for the sub-region, with a maximum score of 100.

[0091] The superpixel segmentation and hierarchical feature evaluation method of this invention enables a precise assessment of the coating uniformity of cylindrical battery electrode sheets. Through gradient calculation in polar coordinate grids and seed point optimization strategies, the superpixel boundaries better conform to the actual coating boundaries. The combination of multi-scale gradient features and local binary mode features comprehensively captures subtle changes on the coating surface. The dynamic fusion weight mechanism adapts to the curvature changes of the cylindrical surface, improving the accuracy of feature representation. The construction of the hierarchical feature system enables multi-scale analysis from local to global perspectives, significantly enhancing the reliability and robustness of coating uniformity evaluation.

[0092] In one alternative implementation,

[0093] The steps of extracting features from the optimized circumferential unfolded image and the axial unfolded image in the polar coordinate domain and the Cartesian coordinate domain, respectively, to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features include:

[0094] An adaptive Gabor filter is constructed in the polar coordinate domain. The radial and angular standard deviations of the adaptive Gabor filter are adaptively adjusted according to the curvature of the cylindrical battery electrode, and the optimized circumferential unfolded image and the axial unfolded image are filtered and enhanced respectively.

[0095] Based on the enhanced optimized circumferential unfolded image and axial unfolded image, a curvature scale space is constructed. The curvature descriptor of the coating edge is calculated under different Gaussian smoothing scales. The curvature descriptor contains the first and second derivatives of the contour point coordinates, and the contour features of the coating edge are obtained.

[0096] The gradients of the enhanced optimized circumferential unfolded image and axial unfolded image are calculated in the Cartesian coordinate domain to construct a structural tensor matrix. The structural tensor matrix is ​​then decomposed into eigenvalues ​​to obtain principal direction eigenvalues. Based on the principal direction eigenvalues, the directional features of the coating surface texture are calculated.

[0097] An adaptive kernel density function is used to estimate the density of the enhanced optimized circumferential and axial unfolded images. The local bandwidth of the adaptive kernel density function is dynamically adjusted according to the gray-level distribution of the coating area to construct density feature vectors at multiple scales and obtain the coating density distribution features.

[0098] In this embodiment, when constructing the adaptive Gabor filter in the polar coordinate domain, the surface curvature characteristics of the cylindrical battery electrode must be considered. For the 18650 cylindrical battery electrode, its surface curvature gradually decreases from the central axis outwards; therefore, the filter parameters should vary with position. The radial standard deviation of the adaptive Gabor filter is set to 3.5 pixels in the paraxial region (0-5mm), 4.0 pixels in the middle region (5-8mm), and 4.5 pixels in the outer region (above 8mm); the angular standard deviation is set to 4.5 degrees in the paraxial region, 3.5 degrees in the middle region, and 2.5 degrees in the outer region. The filter center frequency f0 is set according to the periodicity of the coating texture, typically taking a value of 0.1 pixels - 1. The direction angle θ is selected at 0°, 45°, 90°, and 135° to construct the filter bank. The phase offset ϕ is set to 0° and 90°, corresponding to even-symmetric and odd-symmetric Gabor filters respectively, used to detect linear and edge structures.

[0099] When applying Gabor filter banks to the optimized circumferentially unfolded image in the polar coordinate domain, the image in Cartesian coordinates is first transformed to polar coordinates. During the transformation, the angular resolution is set to 0.5 degrees and the radial resolution to 0.1 mm. A polar coordinate system (r, θ) is established with the cylinder center as the origin, and the gray value of each point in the image is mapped to the polar coordinate plane. For each position (r, θ) in the polar coordinate image, a Gabor filter bank with adaptive parameters is applied for convolution. For a cylindrical battery with a diameter of 18 mm, the image size is typically 2000 × 1000 pixels (circumferential × radial), and the filter window size is set to 25 × 25 pixels. The maximum value of the response in each direction is taken as the enhanced image, improving the clarity of texture edges. The axially unfolded image is also processed by Gabor filtering, but with different parameter settings: the radial standard deviation is uniformly set to 4.0 pixels, and the angular standard deviation is uniformly set to 3.0 degrees to adapt to the characteristics of the axial image.

[0100] After Gabor filtering enhancement, a curvature scale space is constructed to extract the coating edge contour features. Gaussian smoothing scales of 1.0, 2.0, 3.0, and 4.0 pixels are selected sequentially, forming four scale levels. At each scale, Gaussian smoothing is applied to the enhanced image to extract the coating region edges. Edge extraction uses a gray-level thresholding method, with the threshold set to 1.2 times the average gray value of the image. The extracted edges are represented as a set of contour points {(ri,θi)} or {(xi,yi)}, where {(ri,θi)} represents the set of contour points in polar coordinates, where ri represents the radial distance from the i-th point to the central axis of the cylinder, and θi represents the angular position of the i-th point; {(xi,yi)} represents the set of contour points in Cartesian coordinates, where xi represents the horizontal position of the i-th point, and yi represents the vertical position of the i-th point. Contour points are sampled with equal arc lengths at 2mm intervals to ensure the uniformity of the contour representation. Calculate the curvature descriptor for each sampling point, including the first derivative (dr / dθ or dx / dy) and the second derivative (d 2 r / dθ 2 or d 2 x / dy 2 The first derivative is calculated using the central difference method, by dividing the difference between two adjacent points by the sampling interval; the second derivative is calculated by dividing the difference between adjacent first derivatives by the sampling interval.

[0101] For circumferentially unfolded images, the first derivative calculated in polar coordinates reflects the radial rate of change of the edge, while the second derivative reflects the change in edge curvature. The absolute value of the first derivative at a normally coated edge should be less than 0.2, and the absolute value of the second derivative should be less than 0.05; while at defective edges, the absolute value of the first derivative is often greater than 0.5, and the absolute value of the second derivative is greater than 0.1. For axially unfolded images, the derivative is calculated in Cartesian coordinates, with similar criteria. The curvature descriptors at the four scales are concatenated to form a coating edge contour feature vector. Each sampling point in the feature vector contains 8 values ​​(4 scales × 2 derivatives). The number of sampling points is the cell circumference divided by the sampling interval; for an 18650 cell, this is approximately 30 points, therefore the feature vector dimension is 240.

[0102] When calculating texture directionality features in the Cartesian coordinate domain, local region gradient analysis is performed on the enhanced image. The image is divided into overlapping grids of size 16×16 pixels, with adjacent grids overlapping by 8 pixels. For each pixel within a grid, the gradients in the x and y directions are calculated using the Sobel operator: x-direction [-1,0,1;-2,0,2;-1,0,1], y-direction [-1,-2,-1;0,0,0;1,2,1]. A 2×2 structure tensor matrix is ​​constructed based on the gradient values, where the matrix elements are the average of the products of the gradient components within the grid. Eigenvalue decomposition is performed on the structure tensor matrix to obtain two eigenvalues ​​λ1 and λ2 (λ1≥λ2≥0) and their corresponding eigenvectors. The direction corresponding to the principal eigenvector is the principal direction of the local texture, and the eigenvalue ratio λ1 / λ2 reflects the strength of directional consistency. A larger eigenvalue ratio indicates a more pronounced texture directionality; a ratio close to 1 indicates no significant texture directionality. For standard coated areas, the eigenvalue ratio is typically between 2.5 and 4.0; for unevenly coated areas, the eigenvalue ratio is often less than 1.5 or greater than 5.0. The directional consistency strength of each grid is defined as (λ1-λ2) / (λ1+λ2), with a value ranging from [0,1]. A larger value indicates stronger directionality. The directional consistency strength of a normal coated area is typically between 0.5 and 0.7. For all grids in the entire image, a principal orientation angle histogram (dividing 360 degrees into 36 equal intervals) and a directional consistency strength histogram (dividing [0,1] into 10 equal intervals) are calculated. The two histograms are concatenated to form a 46-dimensional directional feature vector of the coated surface texture.

[0103] When using an adaptive kernel density function to extract coating density distribution features, the bandwidth of the kernel function is dynamically adjusted based on the gray-level distribution characteristics of the coating region. The enhanced image is divided into a 32×32 pixel grid. For each grid, the mean μ and standard deviation σ of the gray-level values ​​are calculated. The kernel function bandwidth h is set to be proportional to the standard deviation: h = α·σ, where α is a proportionality coefficient with a value ranging from 0.8 to 1.2. In high-contrast regions (large σ), the bandwidth is increased to smooth noise; in low-contrast regions (small σ), the bandwidth is decreased to preserve details. For example, when the gray-level standard deviation within the grid is 15, the bandwidth is set to 13; when the standard deviation is 30, the bandwidth is set to 30.

[0104] Density estimation is performed on each grid using a Gaussian kernel function, generating a local density distribution map. The kernel function is a two-dimensional Gaussian function in both the spatial and gray-level dimensions. The estimation results are quantized, and the gray-level range [0, 255] is divided into 25 intervals. The density value of each interval is statistically analyzed to form a 25-dimensional local density feature vector. To capture multi-scale characteristics, three grid sizes are designed: 32×32 pixels, 64×64 pixels, and 128×128 pixels, representing micro, meso, and macro scales, respectively. The density feature vectors of the three scales are normalized and concatenated to form a 75-dimensional coating density distribution feature vector. The density distribution in normally coated areas exhibits a unimodal or bimodal shape, with concentrated peak intervals and a small standard deviation; the density distribution in unevenly coated areas often exhibits a multimodal or flat distribution with a large standard deviation.

[0105] Density characteristics can be further described using statistical moments. The mean, standard deviation, skewness, and kurtosis of the density distribution at each scale are calculated, forming a 12-dimensional (3 scales × 4 statistics) statistical feature vector. The mean reflects the overall brightness level, the standard deviation reflects the contrast, the skewness reflects the symmetry of the distribution, and the kurtosis reflects the sharpness of the distribution. In normally coated areas, the absolute value of the skewness is typically less than 0.5, and the kurtosis is close to 3; while in uneven areas, the absolute value of the skewness is often greater than 1.0, and the kurtosis is far from 3. Concatenating the density distribution feature vector with the statistical feature vector forms a complete 87-dimensional coating density distribution feature.

[0106] For severely unevenly coated samples, the absolute value of the second derivative in the edge contour features can reach 0.25, far exceeding the 0.05 of normal samples; in the texture directionality features, the standard deviation of the directionality consistency intensity exceeds 0.3, while normal samples are usually less than 0.15; in the density distribution features, the absolute value of skewness can reach 1.8, and the kurtosis can reach 7.0, significantly deviating from normal samples. These differences in features provide a reliable basis for subsequent evaluation of coating uniformity.

[0107] This invention performs dual-domain feature extraction in both polar and Cartesian coordinates, achieving comprehensive capture of the coating characteristics of cylindrical battery electrodes. An adaptive Gabor filter dynamically adjusts parameters based on curvature changes, improving the accuracy of edge and texture enhancement; curvature scale spatial analysis makes edge contour features more robust; the structural tensor method accurately quantifies texture directionality; and adaptive kernel density estimation accurately characterizes the coating density distribution.

[0108] In one alternative implementation,

[0109] The steps for establishing a coating uniformity evaluation standard include: calculating a coating edge integrity index based on the coating edge contour features, calculating a coating direction consistency index based on the coating surface texture directionality features, and calculating a coating density uniformity index based on the coating density distribution features.

[0110] The coating edge contour features are segmented and fitted, and the curvature continuity score and contour integrity score are calculated to generate a coating edge integrity index.

[0111] Establish a directional feature evaluation criterion, calculate the main direction consistency coefficient of the local area and the global direction distribution entropy value based on the directional features of the coating surface texture, and generate a coating direction consistency index;

[0112] Design a density distribution evaluation function, calculate the regional density mean deviation and density gradient change rate based on the coating density distribution characteristics, and generate a coating density uniformity index.

[0113] The coating edge integrity index, coating direction consistency index, and coating density uniformity index are normalized using a fuzzy comprehensive evaluation method to establish a coating uniformity evaluation standard.

[0114] In this embodiment, the piecewise fitting of the coating edge contour features is achieved using a sliding window approach. For the extracted edge contour point sequence {(ri,θi)} or {(xi,yi)}, a window size of 9 points is set, with adjacent windows overlapping by 5 points. Within each window, piecewise cubic spline interpolation is used to fit the contour points. Specifically, for the point set within each window, a cubic spline curve passing through all points is constructed, ensuring the first and second derivatives of the curve are continuous at the nodes. At the window boundaries, a natural boundary condition is set, i.e., the second derivative is zero. For a cylindrical battery electrode with a diameter of 18mm, the edge contour typically contains approximately 150 sampling points, forming approximately 37 windows. During the fitting process, if the variance of the point set exceeds a threshold (set as 5% of the average contour radius), it indicates drastic contour changes in that region; the window size is then reduced to 5 points to improve fitting accuracy. The curvature continuity score is calculated based on the continuity of derivatives between adjacent fitted segments. For each pair of adjacent window intersections, calculate the absolute values ​​of the first and second derivative differences, d1 and d2, of the difference between the left and right windows. If d1 is less than 0.05 and d2 is less than 0.01, the continuity score is 100 points; if d1 is between 0.05 and 0.1 or d2 is between 0.01 and 0.03, the continuity score is 80 points; if d1 is between 0.1 and 0.2 or d2 is between 0.03 and 0.06, the continuity score is 60 points; if d1 is between 0.2 and 0.3 or d2 is between 0.06 and 0.1, the continuity score is 40 points; if d1 is greater than 0.3 or d2 is greater than 0.1, the continuity score is 20 points. The global curvature continuity score is the weighted average of all intersection score scores, with the weights proportional to the local curvature of the contour.

[0115] The contour integrity score is evaluated based on two aspects: contour closure and contour smoothness. Closure is assessed by the ratio of the distance between the first and last points to the total contour length. A ratio less than 0.5% is considered a complete closure, scoring 100 points; a ratio between 0.5% and 2% is calculated using linear interpolation; and a ratio greater than 2% scores 50 points. Smoothness is assessed by the ratio of the average distance from contour points to the fitted curve to the average contour radius. A ratio less than 1% is considered a complete smoothness, scoring 100 points; a ratio between 1% and 5% is calculated using linear interpolation; and a ratio greater than 5% scores 60 points. The contour integrity score is a weighted average of the closure and smoothness scores, with a weighting ratio of 4:6.

[0116] The coating edge integrity index is obtained by weighting the curvature continuity score and the contour integrity score, with a weighting ratio of 5:5. This index ranges from 0 to 100, with higher values ​​indicating better edge integrity. For standard 18650 cylindrical battery electrodes, the edge integrity index for normally coated areas is typically above 85, for slightly defective areas it is between 70 and 85, and for severely defective areas it is below 70.

[0117] When establishing the directional feature evaluation criterion, the principal direction consistency coefficient of the local region is calculated based on the aforementioned extracted directional features of the coated surface texture. For each 16×16 pixel grid region in the image, its principal direction is the direction of the principal eigenvector obtained by the eigenvalue decomposition of the structure tensor matrix. The principal direction consistency coefficient is defined as (λ1-λ2) / (λ1+λ2+ε), where λ1 and λ2 are eigenvalues ​​(λ1≥λ2≥0), and ε is a small positive number (set to 0.001) to prevent the denominator from being zero. The consistency coefficient ranges from [0,1), and a larger value indicates stronger directionality. For normally coated areas, the principal direction consistency coefficient is usually between 0.6 and 0.8; for unevenly coated areas, the principal direction consistency coefficient is below 0.4 or exhibits a high variance distribution.

[0118] The global directional distribution entropy reflects the degree of disorder in the distribution of main directions throughout the image. The 360-degree area is divided into 36 equal intervals, and the normalized histogram p(i) of the main directions of all grid cells is calculated, i = 1, 2, ..., 36. The information entropy is calculated as the negative sum of all p(i) multiplied by their natural logarithms. The entropy value typically ranges from 0 to 3.6, with smaller values ​​indicating a more concentrated directional distribution. For uniformly oriented coatings, the entropy value is usually less than 2.0; for non-uniformly oriented coatings, the entropy value is usually greater than 2.5. The entropy values ​​are mapped to a scoring scale: entropy values ​​below 1.5 score 100 points, entropy values ​​above 3.0 score 60 points, and intermediate values ​​are obtained through linear interpolation.

[0119] The coating orientation consistency index is jointly determined by the local principal orientation consistency coefficient and the global orientation distribution entropy value. The local index is the average of the principal orientation consistency coefficients of all grids multiplied by 100; the global index is a score based on the entropy value. The coating orientation consistency index is a weighted average of the local and global indices, with a weight ratio of 6:4. This index ranges from 0 to 100, with higher values ​​indicating better orientation consistency. In actual testing, the orientation consistency index of high-quality coated samples is usually above 80.

[0120] When designing the coating density distribution evaluation function, a two-dimensional evaluation model is constructed based on the aforementioned extracted coating density distribution characteristics. This evaluation model includes two key indicators: the deviation of the regional density mean and the change rate of the density gradient. The calculation process of the deviation of the regional density mean is as follows: The image is divided into an 8×8 grid, and each grid corresponds to a physical area of approximately 2mm×2mm on the surface of the cylindrical battery. Calculate the average value μi of the gray values within each grid, where i = 1, 2,..., 64. The global average gray value μg is the mean of all grid average values. The deviation of the regional density mean Dμ is defined as the sum of the absolute values of the differences between the average values of each grid and the global average value, divided by the number of grids, and then divided by the global average value to obtain a normalized deviation value. The scoring function is designed as a piecewise linear function: When Dμ ≤ 0.1, the score Sμ = 100; when 0.1 < Dμ ≤ 0.3, the score Sμ = 100 - 200×(Dμ - 0.1); when Dμ > 0.3, the score Sμ = 60. The Dμ of the ideal coating area is usually within the range of 0.05 - 0.08, corresponding to a score of 100; the Dμ of the slightly uneven area is within the range of 0.15 - 0.25, corresponding to a score of 70 - 90; the Dμ of the severely uneven area can reach above 0.4, corresponding to a score below 60. The deviation of the density mean reflects the uniformity at the macroscopic level and can effectively identify problems of large-area coating omission or accumulation. The density gradient change rate Dg evaluates the smoothness of the gray value change in the local area, and the calculation method is as follows: For each grid, the first-order gradients gx and gy in the x and y directions are calculated using a 3×3 Sobel operator, and the second-order gradients gxx and gyy are calculated using a 5×5 second-order difference template. The gradient change rate is defined as (|gxx| + |gyy|) / (|gx| + |gy| + ε)×10, where ε is a small positive number (taking the value of 0.01) to prevent the denominator from being zero. The scoring function is also designed as a piecewise linear function: When Dg ≤ 0.5, the score Sg = 100; when 0.5 < Dg ≤ 1.5, the score Sg = 100 - 30×(Dg - 0.5); when Dg > 1.5, the score Sg = 70. The density gradient change rate can effectively identify problems of uneven distribution of microscopic coating particles. The final coating density uniformity index S is obtained by a weighted combination of the scores of the deviation of the regional density mean and the density gradient change rate: S = 0.7×Sμ + 0.3×Sg.

[0121] The density gradient change rate evaluates the smoothness of grayscale changes in a local area. For each grid cell, the first-order gradients gx and gy in the x and y directions, and the second-order gradients gxx and gyy are calculated. The gradient change rate is defined as the sum of the absolute values ​​of the second-order gradients divided by the sum of the absolute values ​​of the first-order gradients, multiplied by a constant factor of 10. A gradient change rate less than 0.5 is considered a smooth change and scores 100 points; a change rate between 0.5 and 1.5 is scored using linear interpolation; and a change rate greater than 1.5 scores 70 points. The average gradient change rate score for all grid cells is calculated to obtain the overall density gradient change rate score. The density gradient change rate score for normally coated areas is typically above 85 points.

[0122] The coating density uniformity index is obtained by weighting the regional density mean deviation score and the density gradient change rate score, with a weight ratio of 7:3. This index ranges from 0 to 100, with higher values ​​indicating better density uniformity. For example, products with a coating density uniformity index of 85 or higher are considered high-quality, 70-85 are considered acceptable, and scores below 70 require rework.

[0123] When establishing the coating uniformity evaluation standard using the fuzzy comprehensive evaluation method, the three indicators are first normalized so that they all fall within the range of 0-100 points. The evaluation factor set U={u1,u2,u3} is defined, representing the coating edge integrity index, coating direction consistency index, and coating density uniformity index, respectively. The evaluation level set V={v1,v2,v3,v4,v5} is defined, representing excellent, good, qualified, unqualified, and seriously unqualified, respectively. A membership function μij (i=1,2,3;j=1,2,3,4,5) is established for each indicator ui, representing the membership degree of indicator ui to evaluation level vj.

[0124] The membership functions adopt a trapezoidal distribution: the membership function of the excellent grade is 1 in the interval [90,100] and decreases linearly in the interval [80,90]; the membership function of the good grade is 1 in the interval [80,90] and changes linearly in the intervals [70,80] and [90,95]; the membership function of the qualified grade is 1 in the interval [65,75] and changes linearly in the intervals [55,65] and [75,85]; the membership function of the unqualified grade is 1 in the interval [40,55] and changes linearly in the intervals [30,40] and [55,65]; and the membership function of the seriously unqualified grade is 1 in the interval [0,30] and decreases linearly in the interval [30,40].

[0125] For each indicator ui, its membership degree at each evaluation level is calculated, forming a fuzzy relation matrix R. The weight vector A is determined based on the importance of the indicators; typically, the weight of the edge integrity indicator is set to 0.35, the weight of the directional consistency indicator to 0.3, and the weight of the density uniformity indicator to 0.35. The final evaluation result vector B is obtained through the fuzzy synthesis operation B = A○R, where ○ is the fuzzy synthesis operator, using a weighted average operation. The final evaluation level is determined based on the principle of maximum membership degree.

[0126] To handle boundary conditions in the evaluation, the concept of evaluation confidence is introduced. Confidence is defined as the difference between the largest and second-largest membership degrees, reflecting the reliability of the evaluation result. A confidence score greater than 0.3 is considered a high-confidence evaluation, and less than 0.1 is considered a low-confidence evaluation. For low-confidence evaluation results, a warning message is given, and manual review is recommended.

[0127] This invention establishes a coating uniformity evaluation standard based on coating edge contour features, surface texture directionality features, and density distribution features, enabling an objective quantitative assessment of the coating quality of cylindrical battery electrodes. The fuzzy comprehensive evaluation method integrates multi-dimensional indicators, overcoming the limitations of traditional hard threshold judgments and improving the reliability and stability of the evaluation results.

[0128] In one alternative implementation,

[0129] The steps of constructing a hierarchical feature fusion network, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion, calculating the weight coefficients of circumferential and axial features according to the coating uniformity evaluation criteria, adaptively weighting and fusing the circumferential and axial features, comparing the fusion result with the theoretical coating uniformity distribution, and outputting the coating uniformity detection result include:

[0130] A polar coordinate feature map is constructed from the features extracted in the polar coordinate domain of the optimized circumferential unfolded image, and a Cartesian coordinate feature map is constructed from the features extracted in the Cartesian coordinate domain of the axial unfolded image.

[0131] The polar coordinate attention module constructs a periodic projection matrix based on the angle information in the polar coordinate feature map. The periodic projection matrix has a periodicity of 2π in the angular dimension. The polar coordinate feature map is mapped to a query vector, a key vector, and a value vector through the periodic projection matrix. The product of the query vector and the key vector is calculated, and the product is divided by a preset scaling factor and then passed through a softmax function to obtain the attention weight. The product of the attention weight and the value vector is used as the polar coordinate attention feature.

[0132] The spatial attention module performs convolution, pooling, and multilayer perceptron operations sequentially on the input Cartesian coordinate feature map to obtain spatial attention features.

[0133] Based on the coating uniformity evaluation standard, the circumferential feature weight coefficient and the axial feature weight coefficient are calculated. The circumferential feature weight coefficient and the axial feature weight coefficient are multiplied by the polar coordinate attention feature and the spatial attention feature respectively and then superimposed to obtain the fused feature.

[0134] A theoretical coating uniformity distribution model was established based on coating process parameters;

[0135] Calculate the absolute sum of the differences between the fusion feature and the theoretical coating uniformity distribution model, divide the absolute sum of the differences by the integral value of the theoretical coating uniformity distribution model to obtain the normalized difference degree, and output the coating uniformity detection result based on the normalized difference degree.

[0136] In this embodiment, the construction of the hierarchical feature fusion network begins with the generation of feature maps. The coating edge contour features, coating surface texture directionality features, and coating density distribution features extracted from the optimized circumferentially unfolded image in the polar coordinate domain are integrated into a polar coordinate feature map. Feature channel integration employs dimensionality normalization, mapping each feature vector to the same dimension (typically 128-dimensional) through a fully connected layer, and then concatenating them along the channel dimensions to form a multi-channel feature map. For an 18650 cylindrical battery, the spatial resolution of the circumferentially unfolded image is typically 2048×512 pixels. After feature extraction and downsampling, the resulting polar coordinate feature map has a size of 256×64×384, where 256 represents the angular resolution, 64 represents the radial resolution, and 384 represents the number of feature channels (128 dimensions for each of the three feature classes).

[0137] Features extracted from the axially unfolded image in the Cartesian coordinate domain are also channel-integrated to form a Cartesian coordinate feature map. The spatial resolution of the axially unfolded image is typically 1024×768 pixels, and the Cartesian coordinate feature map obtained after feature extraction and downsampling has a size of 128×96×384. Feature channel normalization is implemented using a batch normalization layer with a mean of 0 and a variance of 1 to ensure numerical comparability between different feature types. For the batch normalization layer, the momentum parameter is set to 0.9 and the epsilon parameter to 10. -5 Statistical parameters are updated during the training phase and fixed during the testing phase.

[0138] The polar coordinate attention module utilizes the periodicity of a cylindrical structure. This module first constructs a periodic projection matrix based on the angle information in the polar coordinate feature map. For each angle θ in the feature map, its periodic function values, including sin(θ) and cos(θ), are calculated to form an angle code. For example, for a feature map with a resolution of 256 angles, an angle range of [0, 2π), and an angle interval of 2π / 256, the sin and cos values ​​at each angle position are calculated, resulting in a periodic encoding matrix of size 256×2. This encoding matrix ensures consistency in feature representation at positions θ and θ+2π, achieving angular periodicity.

[0139] The periodic projection matrix maps the polar coordinate feature map to the query vector, key vector, and value vector through three different linear transformations. The linear transformations are implemented using 1×1 convolutions with 384 input channels and 128 output channels, without using a bias term. For a 256×64×384 polar coordinate feature map, the linear transformations yield query vector Q, key vector K, and value vector V, each with a size of 256×64×128. The scaling factor is set to the square root of the query vector dimension, i.e., 11.3, to normalize the attention score. During attention calculation, the query vector is multiplied by the transpose of the key vector to obtain the original attention score matrix of size 256×64×256×64. To reduce computational complexity, the feature map is flattened in space, resulting in a 16384×128 matrix, and the attention score matrix becomes 16384×16384.

[0140] After dividing the attention score matrix by a scaling factor, a softmax function is applied for normalization to obtain the attention weights. The softmax function is applied independently to each row, ensuring that the sum of the attention weights at each position is 1. The attention weights are multiplied by the value vector to obtain the attention output, which has a size of 16384×128. The attention output is then reconstructed back into a 256×64×128 three-dimensional tensor as the polar coordinate attention feature. In practical applications, to improve computational efficiency, a multi-head attention mechanism is often used, dividing the single 128-dimensional attention head into eight 16-dimensional sub-heads for parallel computation, and finally concatenating the results. Each sub-head uses a different projection matrix to enhance the model's expressive power.

[0141] The spatial attention module processes the Cartesian coordinate feature map, capturing the spatial relationships of axial features. This module first performs convolution operations on the input 128×96×384 feature map. The convolutional layer uses a 3×3 kernel with a stride of 1 and padding of 1, maintaining the spatial dimension and outputting 384 channels. The kernel weights are initialized using the He initialization method, with a standard deviation equal to the reciprocal of the square root of the input channel count, and the bias term is initialized to zero. A batch normalization layer and a ReLU activation function follow the convolution to improve non-linear expressiveness.

[0142] The convolutional feature maps are then pooled using a spatial pyramid pooling strategy. The feature maps are average-pooled using 1×1, 2×2, 4×4, and 8×8 grids to obtain global features at different scales. 1×1 pooling yields a 384-dimensional vector, 2×2 pooling yields a 4×384-dimensional vector, 4×4 pooling yields a 16×384-dimensional vector, and 8×8 pooling yields a 64×384-dimensional vector. These multi-scale features are concatenated, normalized, and then input into a multilayer perceptron. The multilayer perceptron contains two fully connected layers. The first layer has an input dimension of (1+4+16+64)×384=32640 and an output dimension of 1024, using ReLU activation. The second layer has an input dimension of 1024 and an output dimension of 12288 (=128×96), using a sigmoid activation function to restrict the output value to the range of 0-1.

[0143] The output of the multilayer perceptron is reconstructed into a 128×96×1 spatial attention map, which is then multiplied element-wise with the original feature map to obtain a weighted feature. This weighted feature is then passed through a 3×3 convolutional layer with residual connections, outputting a 128×96×128 spatial attention feature. In the residual connections, the convolutional branch contains convolutional layers, batch normalization layers, and ReLU activation functions; the identity mapping branch adjusts the number of channels to 128 through a 1×1 convolution. The outputs of the two branches are added together to form the final spatial attention feature.

[0144] The weighting coefficients for circumferential and axial features are calculated based on the coating uniformity evaluation criteria. The coating uniformity evaluation criteria include edge integrity, directional consistency, and density uniformity. The contributions of the three indicators in the circumferential and axial directions are analyzed: the edge integrity indicator contributes 0.7 in the circumferential direction and 0.3 in the axial direction; the directional consistency indicator contributes 0.4 in the circumferential direction and 0.6 in the axial direction; and the density uniformity indicator contributes 0.5 in both directions. Considering the importance of the three indicators (0.35, 0.3, and 0.35 respectively), the weighting coefficient for the circumferential feature is calculated to be 0.535, and the weighting coefficient for the axial feature is 0.465. The weighting coefficients can be dynamically adjusted through adaptive learning to optimize the fusion process based on sample characteristics.

[0145] The circumferential feature weight coefficients are multiplied by the polar coordinate attention feature, and the axial feature weight coefficients are multiplied by the spatial attention feature to obtain the weighted feature. To fuse the features from the two coordinate systems, coordinate transformation and spatial alignment are required. The polar coordinate attention feature (256×64×128) is transformed to a Cartesian coordinate system size (128×96×128) using bilinear interpolation. During interpolation, the polar coordinate point (r, θ) is mapped to a Cartesian coordinate point (r·cos(θ), r·sin(θ)), and the interpolation result is obtained by finding the four nearest pixels and taking a weighted average. The spatially aligned polar coordinate feature and the spatial attention feature are concatenated along the channel dimension, and then subjected to 1×1 convolution for dimensionality reduction, resulting in an output of 128 channels, yielding the fused feature with a size of 128×96×128.

[0146] The theoretical coating uniformity distribution model is established based on coating process parameters. Key process parameters include coating material viscosity (mPa·s), coating speed (mm / s), coating pressure (kPa), coating temperature (°C), and coating gap (mm). For standard lithium battery electrode coating, typical parameters are: material viscosity 5000 mPa·s, coating speed 20 mm / s, coating pressure 150 kPa, coating temperature 25°C, and coating gap 0.2 mm. Based on fluid dynamics principles, a theoretical model for coating uniformity is established to predict the coating thickness distribution. Under ideal conditions, the coating thickness should satisfy a specific distribution function, gradually increasing from the edge to the center and remaining constant in the central region. For example, for a cylindrical battery with a diameter of 18 mm, the theoretical model predicts a coating thickness of 0.08 mm at the edge (8-9 mm from the center), 0.1 mm in the middle region (4-8 mm from the center), and 0.09 mm in the central region (0-4 mm from the center).

[0147] The theoretical model outputs a 128×96 theoretical coating uniformity distribution map, with the same spatial dimension as the fused features. To calculate the absolute sum of differences between the fused features and the theoretical distribution model, the fused features are first mapped to a single channel via a convolutional layer, resulting in a 128×96×1 actual distribution map. The absolute sum of differences is the sum of the absolute values ​​of the differences between corresponding positions in the actual and theoretical distribution maps. The integral value of the theoretical distribution model is the sum of all elements in the theoretical distribution map. The normalized difference is calculated by dividing the absolute sum of differences by the integral value; the range is typically between 0 and 1, with smaller values ​​indicating a closer approximation to the theoretical model.

[0148] The coating uniformity test results are determined based on normalized difference. Threshold ranges are set: difference less than 0.1 is excellent (level 5), 0.1-0.2 is good (level 4), 0.2-0.3 is acceptable (level 3), 0.3-0.4 is unacceptable (level 2), and greater than 0.4 is seriously unacceptable (level 1). The test results include quality level and a difference heatmap. The heatmap maps the difference between the actual and theoretical distribution maps to a pseudo-color image, visually displaying areas of uneven coating. Red indicates excessive coating thickness, blue indicates excessively thin coating, and green indicates conformity to theoretical expectations. The test results also include defect type identification, classifying defects into types such as irregular edges, thin center, and local clustering based on the heatmap pattern, providing a basis for process improvement.

[0149] The hierarchical feature fusion network constructed in this invention effectively integrates circumferential and axial features. The polar coordinate attention module cleverly handles the periodic continuity of the cylindrical surface through a periodic projection matrix, while the spatial attention module captures the spatial relationships of axial features through multi-scale pooling. Adaptive weight coefficient calculation enables the fusion process to adapt to different sample characteristics, improving feature representation capabilities.

[0150] In one alternative implementation,

[0151] The steps for establishing a theoretical coating uniformity distribution model based on coating process parameters include:

[0152] The coating process parameters include coating slurry viscosity, coating speed, and coating pressure;

[0153] A linear correction relationship between the viscosity of the coating slurry and temperature is established. For a given reference temperature and reference viscosity, a viscosity change coefficient corresponding to a unit temperature change is established to obtain the temperature correction value. Based on the average shear rate in the actual coating process, a viscosity shear thinning correction coefficient table is established, and the shear correction coefficient is obtained by looking up the table. The viscosity correction value is obtained by multiplying the temperature correction value by the shear correction coefficient.

[0154] A piecewise linear function is used to describe the relationship between coating pressure and coating thickness. The coating area is divided into an inlet section, an intermediate section and an outlet section. The thickness correction coefficient of each section is calculated based on the pre-calibrated pressure-thickness correspondence.

[0155] The curvature coefficient is calculated based on the circumferential angle and radial position of the cylindrical battery electrode. The curvature coefficient is then combined with the cosine function of the circumferential angle and the linear function of the radial position to obtain a geometric compensation value.

[0156] The viscosity correction value, thickness correction coefficient, coating speed, and geometric compensation value are combined by weighted averaging to generate a theoretical coating uniformity distribution model.

[0157] In this embodiment, the typical range of coating process parameters during the manufacturing of cylindrical battery electrodes is: slurry viscosity 3000-6000 mPa·s, coating speed 10-30 mm / s, and coating pressure 100-200 kPa. The parameter selection is determined based on the characteristics of the electrode material. For example, lithium-ion battery positive electrode materials typically use a viscosity of 5000 mPa·s, a coating speed of 20 mm / s, and a coating pressure of 150 kPa; negative electrode materials use a viscosity of 4000 mPa·s, a coating speed of 25 mm / s, and a coating pressure of 130 kPa.

[0158] The relationship between viscosity and temperature in coating slurry exhibits a non-linear characteristic, but it can be approximated as linear within a specific temperature range, simplifying calculations. When establishing a linear correction relationship, a reference temperature and viscosity under standard coating conditions are selected as benchmarks. The reference temperature is typically set to 25°C, and the reference viscosity is determined based on the slurry formulation, such as a reference viscosity of 5000 mPa·s for cathode slurry. The linear correction relationship is represented by the viscosity change coefficient k, which is the percentage change in viscosity caused by a 1°C change in temperature. For lithium battery slurries, the k value is typically in the range of 2%-5%, and can be calibrated by measuring viscosity values ​​at different temperatures. For example, for cathode slurry, by measuring viscosities of 5500 mPa·s, 5000 mPa·s, and 4600 mPa·s at 20°C, 25°C, and 30°C respectively, the calculated k value is approximately 3.8%. The temperature correction value is calculated by multiplying the reference viscosity by the temperature deviation and the k value.

[0159] The coating slurry exhibits shear thinning characteristics during the coating process, meaning its viscosity decreases with increasing shear rate. The apparent viscosity of the slurry at different shear rates is measured using a rotational viscometer, and a shear correction factor is calculated. The correction factor table is indexed by shear rate, corresponding to the viscosity ratio at different shear rates. For example, the positive electrode slurry at a shear rate of 50 s... -1 100s -1 300s -1 500s -1 1000s -1 The shear correction factors were 0.85, 0.75, 0.60, 0.50, and 0.40, respectively. In actual coating processes, the average shear rate was estimated based on the coating speed and coating gap, and the corresponding shear correction factor was obtained by referring to a table. For standard coating conditions, when the coating speed was 20 mm / s and the coating gap was 0.1 mm, the average shear rate was approximately 200 s. -1 The corresponding shear correction factor is 0.65. The viscosity correction value is calculated by multiplying the temperature correction value by the shear correction factor. For example, when the temperature correction value is 0.886 and the shear correction factor is 0.65, the viscosity correction value is 0.576, indicating that the viscosity under actual coating conditions is 57.6% of the reference viscosity.

[0160] The relationship between coating pressure and coating thickness is non-linear, but it can be simplified using a piecewise linear function. The coating area is divided into an inlet section, a middle section, and an outlet section, establishing a pressure-thickness correspondence. The inlet section typically comprises the first 15% of the coating area, the outlet section the last 15%, and the middle section the middle 70%. The actual thickness of each section under different pressures is measured through coating experiments to establish a calibration relationship. For example, for cathode coating, when the coating pressure is 100 kPa, 150 kPa, and 200 kPa, the thickness correction factors for the inlet section are 0.95, 0.90, and 0.85, respectively; for the middle section, 1.05, 1.00, and 0.98; and for the outlet section, 1.10, 1.15, and 1.20. In actual coating processes, the thickness correction factors for each section are calculated using linear interpolation based on the set coating pressure. For example, when the coating pressure is 170 kPa, the thickness correction factor for the inlet section is 0.88, the thickness correction factor for the middle section is 0.99, and the thickness correction factor for the outlet section is 1.17.

[0161] The geometric characteristics of cylindrical battery electrodes need to be compensated for using a curvature coefficient. The circumferential angle is defined as the angle from a reference position, ranging from 0 to 360 degrees; the radial position is defined as the distance to the central axis of the cylinder, ranging from 0 to 9 mm for 18650 batteries. The curvature coefficient is defined as 1 / r, where r is the radial position, representing the degree of surface curvature. For a cylindrical battery with a diameter of 18 mm, the outer surface curvature coefficient is 0.111 mm. -1 To avoid the curvature coefficient approaching infinity when the radial position is 0, a modified curvature coefficient of 1 / (r+0.5) can be used, where 0.5mm is the offset. The cosine function of the circumferential angle is used to simulate the periodic variation of the circumferential coating uniformity, with a value range of -1 to 1. The linear function of the radial position is defined as the normalized radial position, i.e., (r-0) / (9-0), with a value range of 0 to 1. The geometric compensation value is calculated by combining the curvature coefficient, cosine function, and linear function. The calculation formula can be expressed as: 1 + 0.1 × curvature coefficient × cosine function + 0.05 × linear function, where 0.1 and 0.05 are weighting coefficients. For example, when the circumferential angle is 90 degrees and the radial position is 6mm, the curvature coefficient is 0.154mm. -1 The cosine function value is 0, the linear function value is 0.667, and the geometric compensation value is 1.033.

[0162] The theoretical coating uniformity distribution model integrates various factors using a weighted average method. The weighted average calculation can be expressed as: w1 × viscosity correction value + w2 × thickness correction coefficient + w3 × coating speed / reference speed + w4 × geometric compensation value, where w1, w2, w3, and w4 are weighting coefficients, and w1 + w2 + w3 + w4 = 1. These weighting coefficients are determined through historical coating data and statistical analysis, reflecting the degree of influence of each factor on coating uniformity. For lithium battery electrode coating, the weighting coefficients are w1 = 0.3, w2 = 0.4, w3 = 0.2, and w4 = 0.1. For example, when the viscosity correction value is 0.576, the inlet section thickness correction factor is 0.88, the coating speed is 25 mm / s, the reference speed is 20 mm / s, and the geometric compensation value is 1.033, the theoretical coating uniformity of the inlet section is 0.3×0.576+0.4×0.88+0.2×1.25+0.1×1.033=0.881.

[0163] For example, for an 18650 battery electrode, the reference temperature is 25℃, the reference viscosity is 5000 mPa·s, and the reference speed is 20 mm / s. When the actual temperature is 28℃, the coating speed is 25 mm / s, and the coating pressure is 170 kPa, the calculated temperature correction value is 0.886, the shear correction factor obtained from the table is 0.65, and the calculated viscosity correction value is 0.576. Based on the coating pressure, the thickness correction factors for the inlet, middle, and outlet sections are calculated to be 0.88, 0.99, and 1.17, respectively. A polar coordinate grid is established for the coating area, and the geometric compensation value for each point is calculated. For example, the center point of the inlet section is located at a circumferential angle of 0 degrees and a radial position of 7.5 mm, and the calculated geometric compensation value is 1.05. By combining all factors through a weighted average, the theoretical coating uniformity at the center point of the inlet section is calculated to be 0.3 × 0.576 + 0.4 × 0.88 + 0.2 × 1.25 + 0.1 × 1.05 = 0.885.

[0164] Another implementation method is also provided: when the viscosity of the coating slurry is expressed as a function of shear rate and temperature, a non-Newtonian fluid model needs to be constructed. The shear rate γ is defined as the slurry velocity gradient, with units of seconds. -1 During the coating process, the shear rate range is typically 10-1000 s. -1Temperature T is expressed in degrees Celsius, and is 25℃ under standard coating conditions. Viscosity η0 under zero shear conditions refers to the viscosity exhibited by the slurry when the shear rate is close to zero, generally obtained by measuring with a low-speed rotational viscometer. The reference temperature T0 is set to 25℃, and the relationship between zero-shear viscosity and temperature is described by the Arrhenius equation. In actual calculations, the relationship between zero-shear viscosity η0(T) and viscosity η0(T0) at the reference temperature T0 can be expressed as: η0(T) = η0(T0)·exp[E / R·(1 / T-1 / T0)], where E is the flow activation energy and R is the gas constant. For lithium battery slurries, a typical E / R value is 3000-4000K. In practical applications, a temperature-dependent zero-shear viscosity model can be established by measuring the zero-shear viscosity at different temperatures (e.g., 15℃, 25℃, 35℃). Taking lithium battery cathode slurry as an example, when the temperature rises from 25°C to 35°C, the zero-shear viscosity decreases from 5000 mPa·s to about 3800 mPa·s; when the temperature drops to 15°C, the zero-shear viscosity increases to about 6800 mPa·s.

[0165] The shear thinning degree is calculated based on the product of the shear rate and the characteristic time. The characteristic time λ represents the structural relaxation time of the slurry, measured in seconds, and is typically determined through stress relaxation experiments. The characteristic time for lithium battery slurries is usually in the range of 0.01-0.1 s. The shear thinning degree α is defined as the dimensionless parameter λγ, representing the degree to which the viscosity of the slurry decreases under shear. The larger the value of α, the more pronounced the shear thinning effect of the slurry. For lithium battery slurries, when the shear rate is 100 s... -1 At this time, the value of α is approximately 1-10. Infinite shear viscosity η∞ refers to the limiting viscosity exhibited by the slurry when the shear rate approaches infinity, typically 10%-30% of the zero shear viscosity. The temperature-corrected viscosity value η is calculated using the Cross model, i.e., η = η∞ + (η0(T) - η∞) / (1 + α) n ), where n is the power law exponent, typically taking a value of 0.8-0.9. For lithium battery slurry, η∞ is usually 500-1500 mPa·s.

[0166] The spatial distribution of the pressure gradient generated by the coating pressure is determined through fluid dynamics analysis. The coating region can be divided into an inlet zone, an intermediate zone, and an outlet zone. The pressure in the inlet zone rises rapidly from ambient pressure to its maximum value, the pressure in the intermediate zone remains relatively stable, and the pressure in the outlet zone drops rapidly to ambient pressure. The pressure gradient dP / dx is defined as the pressure change per unit distance, with units of kPa / mm. In the coating apparatus, the pressure gradient can be obtained by solving the simplified Navier-Stokes equations, considering the rheological properties of the slurry and boundary conditions. For a coating width of 80 mm, when the coating pressure is 150 kPa, the average pressure gradient in the inlet zone (0-15 mm) is approximately 8 kPa / mm, the average pressure gradient in the intermediate zone (15-65 mm) is approximately 0.5 kPa / mm, and the average pressure gradient in the outlet zone (65-80 mm) is approximately -9 kPa / mm. For ease of calculation, a piecewise function can be used to approximate the pressure gradient distribution: inlet zone dP / dx = P max ×[1-exp(-x / L in )]×L in -1 In the intermediate region, dP / dx = k mid ×(xL in ) / (L out -L in ), Export zone dP / dx = -P max ×[1-exp(-(Lx) / L out )]×L out -1 Where: x represents the spatial coordinate in the coating direction, in mm; measured from the coating start position (entry zone); exp represents the natural exponential function, i.e., the power function of e; P max The maximum pressure is L, and the total length of the coating area is L. in and L out k represents the lengths of the entrance and exit areas, respectively. mid The gradient coefficient for the intermediate region is represented by . The spatial distribution of the pressure gradient is discretized into a 128×96 grid, corresponding to the resolution of the coating uniformity detection image.

[0167] The local coating thickness correction value δ is calculated based on the pressure gradient using the Lucas-Washburn equation. When the pressure gradient is positive, the slurry flow accelerates, and the local thickness decreases; when the pressure gradient is negative, the slurry flow decelerates, and the local thickness increases. The correction coefficient k is related to the rheological properties of the slurry and is usually calibrated experimentally. For lithium battery slurries, the k value is approximately 0.05-0.1 mm. 2 / kPa. The local coating thickness correction value is calculated as δ=1+k·(dP / dx), which is dimensionless. Typical correction values ​​are 0.6-0.8 for the inlet region, 0.95-1.05 for the middle region, and 1.2-1.4 for the outlet region. The correction value δ can also be expressed as a Taylor expansion of the pressure gradient: δ=1+k1·(dP / dx)+k2·(dP / dx) 2 Where k1 is a first-order coefficient and k2 is a second-order coefficient. For most lithium battery slurries, k1 is approximately 0.05-0.1 mm. 2 / kPa, k2 is approximately 0.001-0.005mm 4 / kPa 2 The reference thickness h0 refers to the coating thickness under ideal conditions, typically 0.1 mm. The spatial thickness distribution h = h0·δ, in mm.

[0168] The geometric characteristics of cylindrical battery electrodes need to be compensated for using a curvature coefficient. The circumferential angle θ is defined as the angle from a reference position, ranging from 0 to 2π; the radial position r is defined as the distance to the central axis of the cylinder, in mm, and for 18650 batteries, r ranges from 0 to 9 mm. The curvature coefficient κ is defined as 1 / r, in mm. -1 This indicates the degree of surface curvature. For a cylindrical battery with a diameter of 18 mm, the outer surface curvature coefficient is 0.111 mm. -1 The curvature coefficient varies with radial position: at r=3mm, κ=0.333mm. -1 At r=6mm, κ=0.167mm -1 At r=9mm, κ=0.111mm -1 In actual calculations, to avoid the curvature coefficient at r=0 from tending to infinity, a modified curvature coefficient κ'=1 / (r+r0) can be used, where r0 is a small positive number, usually taken as 0.5mm.

[0169] The cosine function cos(θ) of the circumferential angle is used to simulate the periodic variation of circumferential coating uniformity, with a value range of [-1, 1]. This function reflects the periodic thickness variation caused by factors such as equipment vibration and slurry supply fluctuations during the coating process. The linear function f(r) of the radial position is defined as (rr... min ) / (r max -r min ), where r represents the radial distance from the central axis of the cylindrical battery to the current calculation point, in mm; r min For the minimum radial position (usually 0), r maxThe maximum radial position is typically 9 mm. The value of f(r) ranges from [0,1], representing the normalized radial position. This function reflects the linear variation trend of the coating thickness from the inside out, usually due to uneven coating pressure distribution. The geometric compensation value G is calculated as G = 1 + α·κ·cos(θ) + β·f(r), where α and β are weighting coefficients, representing the intensity of the curvature effect and the radial position effect, respectively, typically α = 0.1 and β = 0.05. The values ​​of α and β can be obtained by fitting experimental data using the least squares method. The geometric compensation value ranges from approximately 0.9 to 1.1 and is dimensionless.

[0170] The final calculation of the theoretical coating uniformity distribution model takes all the aforementioned factors into account. The influence of coating speed V on coating uniformity is described by an empirical function, defined as g(V) = 1 + γ·(V-V0) / V0, where V0 is the reference coating speed (usually 20 mm / s), and γ is the speed sensitivity coefficient, with a value of 0.2. The speed sensitivity coefficient γ reflects the degree of influence of coating speed variation on coating thickness and is related to the thixotropic nature of the slurry. When the coating speed is 15 mm / s, g(V) = 0.95; when the coating speed is 25 mm / s, g(V) = 1.05. The speed function g(V) can also be expressed in a more complex form: g(V) = 1 + γ1·(V-V0) / V0 + γ2·[(V-V0) / V0] 2 Where γ1 is a first-order coefficient and γ2 is a second-order coefficient. For lithium battery slurry, γ1 is 0.15-0.25 and γ2 is -0.02 to 0.02. The theoretical coating uniformity distribution model M is calculated as M=g(V)·η·h·G, which represents the theoretical thickness distribution of the coating layer at each location under given process parameters.

[0171] The calculation process of the theoretical model can be decomposed into four main steps: (1) Viscosity correction: calculate the corrected viscosity value η based on temperature and shear rate; (2) Thickness distribution: calculate the spatial thickness distribution h based on pressure gradient; (3) Geometric compensation: calculate the geometric compensation value G based on circumferential angle and radial position; (4) Speed ​​adjustment: calculate the speed function g(V) based on coating speed. Finally, the four factors are multiplied to obtain the final model.

[0172] For example, for an 18650 battery electrode, assuming the coating slurry viscosity is 5000 mPa·s (25℃, zero shear), the coating speed is 20 mm / s, and the coating pressure is 150 kPa, firstly, based on the shear rate during the coating process (approximately 500 s⁻¹),... -1Based on the coating pressure and temperature (25℃), the temperature-corrected viscosity value was calculated to be 2200 mPa·s. When the temperature increased to 30℃, the corrected viscosity value decreased to approximately 2000 mPa·s; when the temperature decreased to 20℃, the corrected viscosity value increased to approximately 2400 mPa·s. Next, the spatial distribution of the pressure gradient was calculated based on the coating pressure to determine the local coating thickness correction value. This correction value was multiplied by the baseline thickness of 0.1 mm to obtain the spatial thickness distribution. For the central region, the thickness was approximately 0.095 mm; for the edge region, the thickness could vary from 0.085 to 0.12 mm. Then, the curvature coefficient and geometric compensation value were calculated based on the circumferential angle and radial position. Finally, the coating speed factor, temperature-corrected viscosity value, spatial thickness distribution, and geometric compensation value were multiplied to obtain a theoretical coating uniformity distribution model with a 128×96 grid.

[0173] During the model validation phase, coating experiments were conducted using different combinations of process parameters. For example, combination one: viscosity 5000 mPa·s, speed 20 mm / s, pressure 150 kPa; combination two: viscosity 4000 mPa·s, speed 25 mm / s, pressure 130 kPa; combination three: viscosity 6000 mPa·s, speed 15 mm / s, pressure 170 kPa. Uniformity testing was performed on the coating samples for each parameter group, and the results were compared with the theoretical model predictions. Comparison methods included average thickness comparison, thickness distribution curve comparison, and thickness uniformity index comparison. To improve the model's adaptability, machine learning methods were used to optimize the model parameters. A large amount of historical coating data was collected, including process parameters and corresponding uniformity test results. The gradient descent algorithm was used to optimize the coefficients in the model, minimizing the error between the theoretical predictions and the actual test results. The optimization objective function can be defined as the mean square error between the predicted and actual values. Parameters were adjusted iteratively until the error converged. The optimized model maintained good predictive performance under different slurry formulations and equipment conditions. The model can be further expanded to incorporate more process parameters, such as coating gap, slurry solids content, and ambient humidity, to improve prediction accuracy.

[0174] The theoretical model can also be combined with an online monitoring system to achieve real-time adjustment of coating parameters. Through a closed-loop control strategy, the difference between actual detection results and theoretical predictions is compared, and coating parameters are dynamically adjusted to maintain optimal coating uniformity. For example, when a region is detected to have insufficient thickness, the system can appropriately reduce the coating speed or increase the coating pressure; when excessive thickness fluctuations are detected, the system can adjust the slurry viscosity or coating gap. This intelligent control method can significantly improve coating uniformity and production efficiency.

[0175] This invention also provides that, on the electrode conveying mechanism of existing testing equipment, multiple LED arrays are added along the electrode movement direction, located upstream, midstream, and downstream of the electrode movement direction, respectively. The upstream position retains the original density / thickness sensor, the midstream position installs a circumferential LED array and a corresponding camera group, and the downstream position installs an axial LED array and a corresponding camera group. The LED arrays are uniformly controlled by the control unit of the existing equipment, and the brightness parameters are adjusted using PWM modulation.

[0176] After the electrode enters the detection area, the sensor collects data on the coating density or thickness to form preliminary reference values. Subsequently, as the electrode passes the mid-position, the circumferential LED array illuminates according to the LED array's emission parameters, and the camera group acquires images from multiple angles. The acquired images are then transformed into polar coordinates to generate a circumferential unfolded image. Similarly, an axial unfolded image is acquired at the downstream position.

[0177] A data fusion interface is established to map the areal density / thickness data measured by the device into a density distribution feature matrix. Simultaneously, coating edge contour features, texture directionality features, and density distribution features are extracted from the image. A weighted average algorithm is used to fuse the density distribution features obtained from image analysis with the directly measured density data, with the weighting coefficients dynamically adjusted based on the standard deviation of the two measurement methods.

[0178] The theoretical coating uniformity distribution model established in this invention combines rheological principles with geometric property analysis, achieving an accurate mapping from process parameters to coating uniformity. The model comprehensively reflects the physical mechanisms involved in the coating process. Through precise calculations of shear thinning degree and temperature correction, combined with geometric compensation of the curvature coefficient, high-precision prediction of coating uniformity is achieved. This model not only provides a theoretical reference for coating uniformity detection but can also be used for process parameter optimization and defect cause analysis, providing a scientific basis for the precise control of the coating process for cylindrical battery electrodes.

[0179] In a second aspect, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

Claims

1. A machine vision-based method for detecting the uniformity of coating on cylindrical battery electrode sheets, characterized in that, include: LED arrays are arranged along the circumferential and axial directions of the cylindrical battery electrode to acquire multi-angle images and unfold the cylindrical surface to obtain circumferential unfolded images and axial unfolded images. Polar coordinate gradient continuity analysis is performed on the circumferential unfolded image to extract the continuity features at the seams. The LED array luminous parameters are adaptively adjusted based on the continuity features at the seams, and the multi-angle images are re-acquired to obtain an optimized circumferential unfolded image. Feature extraction is performed on the optimized circumferential unfolded image and the axial unfolded image in polar coordinate domain and Cartesian coordinate domain respectively to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features. Based on the coating edge contour features, a coating edge integrity index is calculated; based on the coating surface texture directionality features, a coating direction consistency index is calculated; and based on the coating density distribution features, a coating density uniformity index is calculated to establish a coating uniformity evaluation standard. A hierarchical feature fusion network is constructed, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion. The weight coefficients of circumferential and axial features are calculated according to the coating uniformity evaluation criteria. The circumferential and axial features are adaptively weighted and fused. The fusion result is compared with the theoretical coating uniformity distribution, and the coating uniformity detection result is output. The theoretical coating uniformity distribution is calculated based on coating process parameters.

2. The method according to claim 1, characterized in that, The steps of performing polar coordinate gradient continuity analysis on the circumferential unfolded image, extracting continuity features at the seams, adaptively adjusting the LED array luminous parameters based on the continuity features at the seams, and re-acquiring the multi-angle images include: An adaptive polar coordinate mapping relationship is established for the circumferential unfolded image, and the grid density is dynamically adjusted according to the curvature change of the cylindrical surface to obtain a polar coordinate domain image; the theoretical reflection intensity distribution map of the polar coordinate domain image is calculated. The polar coordinate domain image is compared with the theoretical reflection intensity distribution map. The reflection intensity difference map is divided into multiple sub-regions using a superpixel segmentation algorithm. The radial gradient and angular gradient of the sub-regions are calculated. A hierarchical continuity evaluation index is constructed by combining local binary pattern features to obtain the continuity score of each sub-region. A target optimization function is constructed based on the continuity score. The target optimization function includes a weighted sum of seam continuity error, reflection uniformity error, and edge sharpness error. Based on the target optimization function, a particle swarm optimization algorithm is used to iteratively optimize the luminous parameters of the LED array. The particle swarm optimization algorithm is combined with a simulated annealing strategy to dynamically adjust the particle velocity and obtain the initial luminous parameters of the LED array. A self-calibrating marker array is preset in the circumferential unfolded image. The reflection characteristics of the marker array are collected to obtain the environmental change feature vector. A polynomial mapping relationship between the environmental change feature vector and the LED compensation parameters is established, and the real-time LED compensation parameters are calculated. The compensation parameters are applied to the initial light emission parameters to obtain the optimized LED array light emission parameters. The image is re-acquired using the optimized LED array light emission parameters.

3. The method according to claim 2, characterized in that, The steps involved in dividing the reflection intensity difference map into multiple sub-regions using a superpixel segmentation algorithm, calculating the radial and angular gradients of each sub-region, and constructing a hierarchical continuity evaluation index by combining local binary pattern features include: The reflection intensity difference map is preprocessed. A polar coordinate grid is established in the preprocessed reflection intensity difference map based on the surface curvature of the cylindrical battery electrode. The weighted response values ​​of radial gradient and angular gradient are calculated. The seed point position for superpixel segmentation is determined based on the weighted response values. A distance metric function containing spatial distance and color distance terms is constructed with the seed point as the center. Based on the distance metric function, region growing is performed on the preprocessed reflection intensity difference map to obtain multiple sub-regions. Multi-scale radial gradient features are obtained by convolution operation on the sub-region using the multi-scale Sobel operator. The angular gradient features of the sub-region are calculated in polar coordinates. The dynamic fusion weight of the radial gradient features and angular gradient features is determined according to the local curvature of the cylindrical battery electrode to obtain the gradient features of the sub-region. Local binary pattern features are extracted from the sub-region using multiple circular templates with different sampling radii and sampling point numbers. These local binary pattern features are then mapped to rotation-invariant equivalent patterns to obtain the texture features of the sub-region. The area, perimeter, roundness, and eccentricity of the sub-regions are extracted as morphological features; The gradient features, texture features, and morphological features are constructed into a hierarchical feature system, and the continuity score of the sub-region is calculated based on the hierarchical feature system.

4. The method according to claim 1, characterized in that, The steps of extracting features from the optimized circumferential unfolded image and the axial unfolded image in the polar coordinate domain and the Cartesian coordinate domain, respectively, to obtain coating edge contour features, coating surface texture directionality features, and coating density distribution features include: An adaptive Gabor filter is constructed in the polar coordinate domain. The radial and angular standard deviations of the adaptive Gabor filter are adaptively adjusted according to the curvature of the cylindrical battery electrode, and the optimized circumferential unfolded image and the axial unfolded image are filtered and enhanced respectively. Based on the enhanced optimized circumferential unfolded image and axial unfolded image, a curvature scale space is constructed. The curvature descriptor of the coating edge is calculated under different Gaussian smoothing scales. The curvature descriptor contains the first and second derivatives of the contour point coordinates, and the contour features of the coating edge are obtained. The gradients of the enhanced optimized circumferential unfolded image and axial unfolded image are calculated in the Cartesian coordinate domain to construct a structural tensor matrix. The structural tensor matrix is ​​then decomposed into eigenvalues ​​to obtain principal direction eigenvalues. Based on the principal direction eigenvalues, the directional features of the coating surface texture are calculated. An adaptive kernel density function is used to estimate the density of the enhanced optimized circumferential and axial unfolded images. The local bandwidth of the adaptive kernel density function is dynamically adjusted according to the gray-level distribution of the coating area to construct density feature vectors at multiple scales and obtain the coating density distribution features.

5. The method according to claim 1, characterized in that, The steps for establishing a coating uniformity evaluation standard include: calculating a coating edge integrity index based on the coating edge contour features, calculating a coating direction consistency index based on the coating surface texture directionality features, and calculating a coating density uniformity index based on the coating density distribution features. The coating edge contour features are segmented and fitted, and the curvature continuity score and contour integrity score are calculated to generate a coating edge integrity index. Establish a directional feature evaluation criterion, calculate the main direction consistency coefficient of the local area and the global direction distribution entropy value based on the directional features of the coating surface texture, and generate a coating direction consistency index; Design a density distribution evaluation function, calculate the regional density mean deviation and density gradient change rate based on the coating density distribution characteristics, and generate a coating density uniformity index. The coating edge integrity index, coating direction consistency index, and coating density uniformity index are normalized using a fuzzy comprehensive evaluation method to establish a coating uniformity evaluation standard.

6. The method according to claim 1, characterized in that, The steps of constructing a hierarchical feature fusion network, including a polar coordinate attention module for circumferential feature fusion and a spatial attention module for axial feature fusion, calculating the weight coefficients of circumferential and axial features according to the coating uniformity evaluation criteria, adaptively weighting and fusing the circumferential and axial features, comparing the fusion result with the theoretical coating uniformity distribution, and outputting the coating uniformity detection result include: A polar coordinate feature map is constructed from the features extracted in the polar coordinate domain of the optimized circumferential unfolded image, and a Cartesian coordinate feature map is constructed from the features extracted in the Cartesian coordinate domain of the axial unfolded image. The polar coordinate attention module constructs a periodic projection matrix based on the angle information in the polar coordinate feature map. The periodic projection matrix maps the polar coordinate feature map to a query vector, a key vector, and a value vector. The product of the query vector and the key vector is calculated, divided by a preset scaling factor, and then passed through a softmax function to obtain the attention weight. The product of the attention weight and the value vector is used as the polar coordinate attention feature. The spatial attention module performs convolution, pooling, and multilayer perceptron operations sequentially on the input Cartesian coordinate feature map to obtain spatial attention features. Based on the coating uniformity evaluation standard, the circumferential feature weight coefficient and the axial feature weight coefficient are calculated. The circumferential feature weight coefficient and the axial feature weight coefficient are multiplied by the polar coordinate attention feature and the spatial attention feature respectively and then superimposed to obtain the fused feature. A theoretical coating uniformity distribution model was established based on coating process parameters; Calculate the absolute sum of the differences between the fusion feature and the theoretical coating uniformity distribution model, divide the absolute sum of the differences by the integral value of the theoretical coating uniformity distribution model to obtain the normalized difference degree, and output the coating uniformity detection result based on the normalized difference degree.

7. The method according to claim 6, characterized in that, The steps for establishing a theoretical coating uniformity distribution model based on coating process parameters include: The coating process parameters include coating slurry viscosity, coating speed, and coating pressure; A linear correction relationship between the viscosity of the coating slurry and temperature is established. For a given reference temperature and reference viscosity, a viscosity change coefficient corresponding to a unit temperature change is established to obtain the temperature correction value. Based on the average shear rate in the actual coating process, a viscosity shear thinning correction coefficient table is established, and the shear correction coefficient is obtained by looking up the table. The viscosity correction value is obtained by multiplying the temperature correction value by the shear correction coefficient. A piecewise linear function is used to describe the relationship between coating pressure and coating thickness. The coating area is divided into an inlet section, an intermediate section and an outlet section. The thickness correction coefficient of each section is calculated based on the pre-calibrated pressure-thickness correspondence. The curvature coefficient is calculated based on the circumferential angle and radial position of the cylindrical battery electrode. The curvature coefficient is then combined with the cosine function of the circumferential angle and the linear function of the radial position to obtain a geometric compensation value. The viscosity correction value, thickness correction coefficient, coating speed, and geometric compensation value are combined by weighted averaging to generate a theoretical coating uniformity distribution model.

Images