Grinding wheel tooth edge visual recognition and wear degree evaluation algorithm

Abstract

This invention relates to an algorithm for visual recognition and wear assessment of the edge of a grinding wheel tooth in a grinding machine, belonging to the field of advanced manufacturing and automation technology. The algorithm includes the following steps: S1: Image acquisition and region of interest extraction; S2: Image preprocessing and enhancement; S3: Robust localization and anomaly identification of the grinding wheel tooth edge; S4: Extraction of wear feature parameters of the grinding wheel tooth; S5: Comprehensive assessment and grading of wear degree. This invention solves the problems of low accuracy in edge recognition and inaccurate wear assessment of grinding wheel teeth in grinding machines. By using sub-pixel interpolation technology, it overcomes the physical limitations of hardware resolution, improving measurement accuracy. Addressing the characteristics of easily broken and blurred wear edges, a confidence level division and RANSAC repair mechanism are designed to effectively distinguish normal contours from wear defects and avoid noise interference. A multi-dimensional feature vector containing deviation, curvature, area, and roughness is constructed, and a machine learning model is used for comprehensive evaluation, resulting in more objective and accurate results.

Description

Technical Field

[0001] This invention relates to the field of advanced manufacturing and automation technology, specifically to an algorithm for visual recognition of the tooth edges of a grinding wheel and assessment of wear degree. Background Technology

[0002] Wafer thinning technology is a critical process step in semiconductor manufacturing. As semiconductor devices evolve towards higher integration and three-dimensional stacking, increasingly stringent requirements are being placed on the thickness uniformity and surface quality of the back-side thinning of wafers. Wafer thinning and polishing machines use high-speed rotating grinding wheels to remove material from the back side of the wafer. The wear condition of the grinding wheel teeth, as the directly involved working part in the grinding process, directly affects the quality and efficiency of wafer thinning. When the grinding wheel teeth wear, chip, or experience material adhesion, it can lead to serious defects on the wafer surface such as scratches, uneven thickness, or even wafer breakage. Therefore, online monitoring and accurate assessment of the wear degree of the grinding wheel teeth have become crucial technical requirements for ensuring the stability of the wafer thinning process.

[0003] A patent publication (CN221774262U) discloses an online damage detection device for a thinning machine grinding wheel. This device includes at least one set of through-beam sensors and an adjustment mechanism. The through-beam sensors include a transmitter and a receiver, respectively positioned on opposite sides of the grinding wheel. Light emitted from the transmitter is received by the receiver after passing through the grinding wheel. The grinding wheel may cut off the transmitted light or change the amount of light transmitted. Online detection is achieved by detecting changes in the amount of light transmitted through the through-beam sensors. This technical solution can determine whether the grinding wheel has suffered large-area damage or loss by detecting changes in light transmission, providing a feasible technical approach for online monitoring of the grinding wheel's condition. However, this technical solution has significant limitations in practical applications. Its through-beam sensors can only detect whether the light is blocked or whether the overall amount of light transmitted changes. It is difficult to detect minute wear and localized chipping on the edges of the grinding wheel teeth. Especially when micron-level edge damage occurs on the grinding wheel teeth, the change in light transmission is extremely weak and difficult to effectively identify, resulting in detection accuracy that cannot meet the requirements of precision thinning processes. Furthermore, this solution cannot quantify the wear of the grinding wheel teeth, and can only provide a binary judgment of whether they are damaged or not. It lacks effective monitoring methods for the progressive wear process, making it impossible for operators to predict the remaining life of the grinding wheel based on the wear trend, and making it difficult to arrange maintenance plans in a timely manner to avoid sudden failures.

[0004] Further research revealed existing technologies that employ visual recognition technology to detect wafer edge chipping. For example, patent application CN118720867A involves a device for targeted removal of wafer edge chipping, which uses a visual recognition component to capture wafer images and identify the chipping width at various circumferential positions. While this approach introduces visual technology into wafer processing, it detects edge chipping defects on the wafer itself rather than the wear state of the grinding wheel teeth. Furthermore, its technical solution only focuses on identifying macroscopic defects at the wafer edge, without addressing high-precision positioning and wear quantification of the grinding wheel tooth edges. At the image processing level, existing visual inspection methods typically use conventional edge detection operators for pixel-level positioning. When processing micro-chipping areas caused by wear on the grinding wheel teeth, the presence of abrupt changes in edge curvature leads to Runge's phenomenon in traditional grayscale interpolation methods, causing oscillations in the fitted curve. This results in multiple pseudo-edge points being identified as real edges, causing edge topology confusion. The algorithm cannot distinguish which edge points represent the macroscopic contour of the teeth and which represent abnormal edges caused by chipping. This technical defect makes it easy for existing solutions to misjudge severe chipping as ordinary wear in practical applications, missing the best maintenance opportunity, which in turn leads to abnormal contact between the grinding wheel and the wafer, resulting in serious consequences such as over-grinding and scrapping of the wafer or even breakage of the grinding wheel. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a visual recognition and wear assessment algorithm for the edges of teeth on grinding wheels. Through sub-pixel edge localization, outlier repair, and multi-dimensional feature fusion, it solves the problems of low accuracy in edge recognition and inaccurate wear assessment of grinding wheel teeth. Sub-pixel interpolation technology overcomes the physical limitations of hardware resolution, improving measurement accuracy. Addressing the characteristics of easily broken and blurred wear edges, a confidence level division and RANSAC repair mechanism are designed to effectively distinguish normal contours from wear defects and avoid noise interference. Abandoning a single judgment criterion, a multi-dimensional feature vector including deviation, curvature, area, and roughness is constructed, and a machine learning model is used for comprehensive evaluation, resulting in more objective and accurate results.

[0006] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: An algorithm for visual recognition and wear assessment of the tooth edges of a grinding wheel in a grinding machine includes the following steps: S1: Image acquisition and region of interest extraction. The original image of the grinding wheel teeth of the grinding machine is acquired by an industrial camera under preset lighting conditions. Based on the geometric arrangement features of the grinding wheel teeth and a preset template matching algorithm, the region of interest image containing one or more grinding wheel teeth is accurately extracted from the original image and used as the input image for subsequent processing.

[0007] S2: Image preprocessing and enhancement. The image of the region of interest is converted to grayscale, and a multi-scale noise reduction algorithm based on wavelet transform is used to remove noise during the image acquisition process. Then, an adaptive contrast enhancement method based on histogram equalization is applied to highlight the grayscale difference between the edge of the grinding wheel teeth and the background, and the enhanced grinding wheel teeth image is obtained.

[0008] S3: Robust localization and anomaly identification of grinding wheel tooth edges. After receiving the enhanced grinding wheel tooth image, the Canny edge detection operator is first used to perform pixel-level coarse edge localization to obtain an initial edge point set. Then, for each edge point in the initial edge point set, the consistency of its local gradient direction is calculated to generate an edge confidence map. Based on the preset confidence threshold, the initial edge point set is divided into a smooth region point set and a high-risk anomaly region point set.

[0009] For a smooth region point set, cubic spline interpolation is performed on the gray values ​​along the gradient direction to reconstruct the gray-level distribution surface in the neighborhood of the point. The sub-pixel coordinates of the edge are accurately located by solving for the points where the second derivative of the surface is zero, thus obtaining the first sub-pixel contour point set.

[0010] For the high-risk anomaly point set, a robust fitting method based on the random sampling consensus algorithm is adopted. Through iterative random sampling and model verification, a low-order spline curve representing the macroscopic geometric trend of the grinding wheel teeth is fitted. Using this curve as a benchmark, points that deviate from the curve by more than a preset distance threshold are marked as anomaly point sets. At the same time, the part of the curve corresponding to the high-risk anomaly area is sub-pixel encrypted to obtain the second sub-pixel contour point set.

[0011] Finally, the first subpixel contour point set and the second subpixel contour point set are merged to generate the repaired subpixel-level macroscopic contour point set, and the abnormal point set is output separately.

[0012] S4: Extraction of wear feature parameters of grinding wheel teeth. Receive the repaired sub-pixel level macroscopic contour point set and the abnormal point set. Based on the repaired sub-pixel level macroscopic contour point set, fit the standard geometric contour of the grinding wheel teeth using the least squares method, and calculate the deviation between the actual contour and the standard contour. Simultaneously, extract the local curvature change and wear area of ​​the grinding wheel tooth edge based on the abnormal point set, and calculate the edge roughness based on the repaired sub-pixel level macroscopic contour point set. Use the deviation, local curvature change, wear area, and edge roughness together as feature parameters to quantify the degree of wear.

[0013] S5: Comprehensive assessment and grading of wear degree. It receives the deviation, local curvature change, wear area and edge roughness extracted in step S4 and inputs them into a pre-built support vector machine-based wear degree assessment model. This model is trained with historical wear data and outputs the wear degree level of the grinding wheel teeth. The wear degree level includes slight wear, moderate wear and severe wear.

[0014] Preferably, in step S1, image acquisition and region of interest extraction use blue monochromatic light with a wavelength of 450nm to 460nm as a preset light source to eliminate ambient light interference and enhance the contrast of the grinding wheel tooth edges; the industrial camera has a resolution of more than 5 million pixels and acquires the original image when the grinding wheel is stationary; the template matching algorithm calculates the similarity between the original image and the pre-stored standard tooth template based on the normalized cross-correlation function, and the region corresponding to the peak similarity position is the region of interest, which is cropped from the original image and rotated to the standard pose before being output to step S2.

[0015] Preferably, the wavelet transform-based multi-scale denoising algorithm in step S2 converts the region of interest image into a grayscale image, and then performs three-level two-dimensional discrete wavelet decomposition using the Daubechies-4 wavelet basis to obtain the low-frequency approximate components. And high-frequency detail components in the horizontal, vertical and diagonal directions. ,in Indicates the number of decomposition layers; applies a soft thresholding function to shrink the coefficients of the high-frequency detail components in each direction of each layer. ;in These are the original wavelet coefficients. The threshold adjustment factor is set to 2.5. The noise standard deviation estimate is calculated using the median absolute deviation method; finally, the denoised image is reconstructed through wavelet inverse transform and output to the subsequent histogram equalization process.

[0016] Preferably, the adaptive contrast enhancement method in step S2 involves receiving the wavelet-denoised image, dividing it into several non-overlapping local sub-blocks, each sub-block being 64×64 pixels in size; calculating the grayscale histogram for each sub-block separately, and setting a cropping limiting coefficient. The gray levels in the histogram that are higher than the total number of pixels in the sub-blocks are cropped and the cropped parts are evenly redistributed to all gray levels to limit contrast amplification. Then, histogram equalization is performed on each sub-block, and bilinear interpolation is used to eliminate the block effect at the sub-block boundary. Finally, the contrast-enhanced grinding wheel tooth image is output to step S3.

[0017] Preferably, the specific process of generating the edge confidence map and dividing the point set in step S3 is as follows: receiving the enhanced image output from step S2, firstly, using the Canny operator to extract pixel-level edge points to obtain the initial edge point set. For each edge point Take a 5×5 neighborhood window centered on it, and calculate the gradient direction angle of all pixels within the window. ,in Given the pixel coordinates within the neighborhood; then calculate the variance of the gradient direction angle. in The mean of the gradient direction angles within the neighborhood; the variance The reciprocal of the normalized value is used as the confidence level for that point. Generate an edge confidence map; preset confidence threshold. ,Will The points are assigned to the smooth region point set. ,Will The points are classified into the high-risk anomaly point set. The results are then output to subsequent processing.

[0018] Preferably, in step S3, the smoothed region point set is received. For each of the edge points Cubic spline interpolation is performed along the gradient direction with a step size of 0.15 pixels to reconstruct the gray-level distribution function along the normal direction of the point. ,in Let the coordinates be along the normal direction; solve the equation. Obtain the precise location of the grayscale inflection point Mapping it back to the image coordinate system yields subpixel coordinates. For all Perform the above operations on the points in the matrix to generate the first sub-pixel contour point set. The output is then fed into subsequent fusion steps to complete the sub-pixel localization of the point set in the smooth region.

[0019] Preferably, the robust fitting process for the high-risk anomaly point set in step S3 is to receive the high-risk anomaly point set. A robust fitting method based on the random sampling consensus algorithm is adopted, and a maximum number of iterations is set. Interior point distance threshold Pixels; in each iteration, from Three points are randomly selected, and a second-order spline curve is fitted as a candidate model. ;calculate All points The Euclidean distance, the statistical distance is less than The points are taken as interior points, and the number of interior points is recorded.

[0020] After the iteration is complete, the candidate model with the most interior points is selected as the optimal model. That is, a low-order spline curve characterizing the macroscopic geometric trend of the grinding wheel teeth; medium to The distance is greater than The points are marked as anomaly point set At the same time exist The corresponding intervals are sampled with encrypted sampling at 0.1 pixel intervals to obtain the second sub-pixel contour point set. ; final fusion and Generate the repaired subpixel-level macroscopic contour point set and will Output separately to step S4.

[0021] Preferably, the wear feature parameter extraction in step S4 specifically includes: receiving the repaired sub-pixel level macroscopic contour point set. and anomaly set ;based on The standard geometric profile is fitted using the least squares method. Let the equation of the standard profile be... ,for Each point in Calculate its directed distance to the standard profile: ; Recalculate the deviation characteristics: ; Local curvature variation characteristics Through calculation The standard deviation of the curvature of the arc formed by every three adjacent points is obtained; the area characteristics of the wear zone are also described. Through calculation The total pixel area of ​​the connected component formed by all outliers is obtained by specifically... Perform convex hull detection and calculate the convex hull area; edge roughness features. based on Calculation, in For each point in the standard profile, take a window of length 11 points along the edge normal. Calculate the root mean square (RMS) distance from all points within the window to the standard profile, and then average the RMS distances of all points. As a four-dimensional feature vector, it is output to step S5.

[0022] Preferably, the support vector machine model construction and training in step S5 specifically includes: pre-collecting 500 historical grinding wheel tooth samples each of three categories: slight wear, moderate wear, and severe wear; and extracting the feature vector of each sample through steps S1 to S4. The training dataset is constructed; Z-score normalization is performed on each feature dimension, i.e. ,in and These are the mean and standard deviation of the feature on the training set, respectively; a radial basis function is used. Construct a one-to-many support vector machine multi-classifier and optimize the penalty parameter using a grid search method. and kernel parameters Finally, the optimal model is obtained; in the evaluation phase, the four-dimensional feature vector output from step S4 is received, and after the same normalization, it is input into the model. The model outputs the decision value for each category, and the category corresponding to the maximum value is taken as the wear level.

[0023] The present invention can achieve the following effects: This invention provides an algorithm for visual recognition and wear assessment of tooth edges on grinding wheels of a grinding machine. Compared with existing technologies, it solves the problems of low accuracy in tooth edge recognition and inaccurate wear assessment by using sub-pixel edge localization, outlier repair, and multi-dimensional feature fusion. It has the following technical effects: I. A region segmentation method based on edge confidence maps: This method first uses the Canny operator to extract pixel-level edge points of the grinding wheel teeth, then calculates the variance of the gradient direction angle in the neighborhood of each edge point, generates an edge confidence map based on the variance, and divides the initial edge point set into a smooth region set and a high-risk abnormal region set based on a preset confidence threshold. This achieves intelligent classification processing of grinding wheel tooth edges, accurately identifying regions with abrupt changes in edge curvature caused by wear and chipping. It avoids the edge topology confusion problem caused by mixing these abnormal regions with smooth regions using the same interpolation algorithm in subsequent processing, laying a reliable data foundation for robust localization of grinding wheel tooth edges.

[0024] II. A robust fitting method based on the random sampling consensus algorithm. This method targets high-risk anomaly point sets and, through iterative random sampling and model validation mechanisms, fits a low-order spline curve representing the macroscopic geometric trend of the grinding wheel teeth. Points deviating from this curve are marked as anomalies. Simultaneously, sub-pixel densification is performed on the corresponding regions on the curve to obtain the repaired contour point set. This overcomes the defect of traditional cubic spline interpolation, which produces Runge phenomenon at abrupt changes in edge curvature, causing oscillations in the fitted curve. It can effectively restore the true macroscopic contour of the grinding wheel teeth from severely chipped or material-adhesive localized damage areas, ensuring that sub-pixel positioning accuracy is not affected by local anomalies.

[0025] Third, a source-specific extraction method for wear feature parameters is proposed. This method uses the repaired sub-pixel-level macroscopic contour point set to calculate the overall deviation and edge roughness, while using the separately output anomaly point set to calculate local curvature changes and wear area. This achieves specialized processing of data from different sources, avoiding the introduction of abnormal edge points generated by chipping into macroscopic contour fitting and overall deviation calculation, thereby preventing edge topology confusion and misjudgment caused by interpolation oscillations. At the same time, it fully utilizes the anomaly point set to accurately quantify the local features of the wear area, making the wear degree assessment more comprehensive and accurate.

[0026] IV. A method for assessing the wear degree of grinding wheel teeth is proposed, integrating edge confidence region segmentation, robust fitting repair, and source feature extraction. This method identifies high-risk areas using a confidence map and performs macroscopic contour repair using a random sampling consensus algorithm. The repaired contour point set and outlier point set are used to calculate different types of wear feature parameters. Finally, the deviation, local curvature change, wear area, and edge roughness are input into a support vector machine model to determine the wear level. A complete data processing chain from image acquisition to wear grading assessment is constructed, solving the technical problem of positioning errors easily generated at abrupt changes in edge curvature in traditional visual inspection methods. This achieves accurate quantification and intelligent grading of the wear degree of grinding wheel teeth, providing reliable technical support for quality control in wafer thinning processes. Attached Figure Description

[0027] Figure 1 This is the overall flowchart of the present invention.

[0028] Figure 2 This is a flowchart of the robust positioning and abnormal point identification of the grinding wheel tooth edge of the present invention.

[0029] Figure 3 This is a flowchart of the wear characteristic parameter extraction process of the present invention. Detailed Implementation

[0030] The technical solution of the invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0031] Example: Figure 1 , Figure 2 and Figure 3 As shown, an algorithm for visual recognition of the tooth edge and assessment of wear degree of a grinding wheel in a grinding machine includes the following steps: S1: Image acquisition and region of interest extraction. The original image of the grinding wheel teeth of the grinding machine is acquired by an industrial camera under preset lighting conditions. Based on the geometric arrangement features of the grinding wheel teeth and a preset template matching algorithm, the region of interest image containing one or more grinding wheel teeth is accurately extracted from the original image and used as the input image for subsequent processing.

[0032] Image acquisition and region of interest extraction use blue monochromatic light with a wavelength of 450nm to 460nm as a preset light source to eliminate ambient light interference and enhance the contrast of the grinding wheel tooth edges; the industrial camera has a resolution of more than 5 million pixels and acquires the original image when the grinding wheel is stationary; the template matching algorithm calculates the similarity between the original image and the pre-stored standard tooth template based on the normalized cross-correlation function, and the region corresponding to the similarity peak position is the region of interest, which is cropped from the original image and rotated to the standard pose before being output to step S2.

[0033] S2: Image preprocessing and enhancement. The image of the region of interest is converted to grayscale, and a multi-scale noise reduction algorithm based on wavelet transform is used to remove noise during the image acquisition process. Then, an adaptive contrast enhancement method based on histogram equalization is applied to highlight the grayscale difference between the edge of the grinding wheel teeth and the background, and the enhanced grinding wheel teeth image is obtained.

[0034] The multi-scale noise reduction algorithm based on wavelet transform converts the region of interest image into a grayscale image, and then performs three-level two-dimensional discrete wavelet decomposition using the Daubechies-4 wavelet basis to obtain the low-frequency approximate components. And high-frequency detail components in the horizontal, vertical and diagonal directions. ,in Indicates the number of decomposition layers; applies a soft thresholding function to shrink the coefficients of the high-frequency detail components in each direction of each layer. ;in These are the original wavelet coefficients. The threshold adjustment factor is set to 2.5. The noise standard deviation estimate is calculated using the median absolute deviation method; finally, the denoised image is reconstructed through wavelet inverse transform and output to the subsequent histogram equalization process.

[0035] The adaptive contrast enhancement method specifically involves receiving an image after wavelet denoising, dividing it into several non-overlapping local sub-blocks, each sub-block being 64×64 pixels in size; calculating the gray-level histogram for each sub-block separately, and setting a cropping limiting coefficient. The gray levels in the histogram that are higher than the total number of pixels in the sub-blocks are cropped and the cropped parts are evenly redistributed to all gray levels to limit contrast amplification. Then, histogram equalization is performed on each sub-block, and bilinear interpolation is used to eliminate the block effect at the sub-block boundary. Finally, the contrast-enhanced grinding wheel tooth image is output to step S3.

[0036] S3: Robust localization and anomaly identification of grinding wheel tooth edges. After receiving the enhanced grinding wheel tooth image, the Canny edge detection operator is first used to perform pixel-level coarse edge localization to obtain an initial edge point set. Then, for each edge point in the initial edge point set, the consistency of its local gradient direction is calculated to generate an edge confidence map. Based on the preset confidence threshold, the initial edge point set is divided into a smooth region point set and a high-risk anomaly region point set.

[0037] The specific process of generating the edge confidence map and dividing the point set is as follows: receiving the enhanced image output from step S2, firstly, the Canny operator is used to extract pixel-level edge points to obtain the initial edge point set. For each edge point Take a 5×5 neighborhood window centered on it, and calculate the gradient direction angle of all pixels within the window. ,in Given the pixel coordinates within the neighborhood; then calculate the variance of the gradient direction angle. in The mean of the gradient direction angles within the neighborhood; the variance The reciprocal of the normalized value is used as the confidence level for that point. Generate an edge confidence map; preset confidence threshold. ,Will The points are assigned to the smooth region point set. ,Will The points are classified into the high-risk anomaly point set. The results are then output to subsequent processing.

[0038] For a smooth region point set, cubic spline interpolation is performed on the gray values ​​along the gradient direction to reconstruct the gray-level distribution surface in the neighborhood of the point. The sub-pixel coordinates of the edge are accurately located by solving for the points where the second derivative of the surface is zero, thus obtaining the first sub-pixel contour point set.

[0039] By receiving the set of points in the smooth region For each of the edge points Cubic spline interpolation is performed along the gradient direction with a step size of 0.15 pixels to reconstruct the gray-level distribution function along the normal direction of the point. ,in Let the coordinates be along the normal direction; solve the equation. Obtain the precise location of the grayscale inflection point Mapping it back to the image coordinate system yields subpixel coordinates. For all Perform the above operations on the points in the matrix to generate the first sub-pixel contour point set. The output is then fed into subsequent fusion steps to complete the sub-pixel localization of the point set in the smooth region.

[0040] For the high-risk anomaly point set, a robust fitting method based on the random sampling consensus algorithm is adopted. Through iterative random sampling and model verification, a low-order spline curve representing the macroscopic geometric trend of the grinding wheel teeth is fitted. Using this curve as a benchmark, points that deviate from the curve by more than a preset distance threshold are marked as anomaly point sets. At the same time, the part of the curve corresponding to the high-risk anomaly area is sub-pixel encrypted to obtain the second sub-pixel contour point set.

[0041] The robust fitting process for the high-risk anomaly point set involves receiving the high-risk anomaly point set. A robust fitting method based on the random sampling consensus algorithm is adopted, and a maximum number of iterations is set. Interior point distance threshold Pixels; in each iteration, from Three points are randomly selected, and a second-order spline curve is fitted as a candidate model. ;calculate All points The Euclidean distance, the statistical distance is less than The points are taken as interior points, and the number of interior points is recorded.

[0042] After the iteration is complete, the candidate model with the most interior points is selected as the optimal model. That is, a low-order spline curve characterizing the macroscopic geometric trend of the grinding wheel teeth; medium to The distance is greater than The points are marked as anomaly point set At the same time exist The corresponding intervals are sampled with encrypted sampling at 0.1 pixel intervals to obtain the second sub-pixel contour point set. ; final fusion and Generate the repaired subpixel-level macroscopic contour point set and will Output separately to step S4.

[0043] Finally, the first subpixel contour point set and the second subpixel contour point set are merged to generate the repaired subpixel-level macroscopic contour point set, and the abnormal point set is output separately.

[0044] S4: Extraction of wear feature parameters of grinding wheel teeth. Receive the repaired sub-pixel level macroscopic contour point set and the abnormal point set. Based on the repaired sub-pixel level macroscopic contour point set, fit the standard geometric contour of the grinding wheel teeth using the least squares method, and calculate the deviation between the actual contour and the standard contour. Simultaneously, extract the local curvature change and wear area of ​​the grinding wheel tooth edge based on the abnormal point set, and calculate the edge roughness based on the repaired sub-pixel level macroscopic contour point set. Use the deviation, local curvature change, wear area, and edge roughness together as feature parameters to quantify the degree of wear.

[0045] Wear feature parameter extraction specifically includes: receiving the repaired sub-pixel level macroscopic contour point set. and anomaly set ;based on The standard geometric profile is fitted using the least squares method. Let the equation of the standard profile be... ,for Each point in Calculate its directed distance to the standard profile: ; Recalculate the deviation characteristics: ; Local curvature variation characteristics Through calculation The standard deviation of the curvature of the arc formed by every three adjacent points is obtained; the area characteristics of the wear zone are also described. Through calculation The total pixel area of ​​the connected component formed by all outliers is obtained by specifically... Perform convex hull detection and calculate the convex hull area; edge roughness features. based on Calculation, in For each point in the standard profile, take a window of length 11 points along the edge normal. Calculate the root mean square (RMS) distance from all points within the window to the standard profile, and then average the RMS distances of all points to obtain the final profile. ;Will As a four-dimensional feature vector, it is output to step S5.

[0046] S5: Comprehensive assessment and grading of wear degree. It receives the deviation, local curvature change, wear area and edge roughness extracted in step S4 and inputs them into a pre-built support vector machine-based wear degree assessment model. This model is trained with historical wear data and outputs the wear degree level of the grinding wheel teeth. The wear degree level includes slight wear, moderate wear and severe wear.

[0047] The support vector machine model construction and training specifically includes: pre-collecting 500 historical grinding wheel tooth samples each of three categories: slight wear, moderate wear, and severe wear; and extracting the feature vector of each sample through steps S1 to S4. The training dataset is constructed; Z-score normalization is performed on each feature dimension, i.e. ,in and These are the mean and standard deviation of the feature on the training set, respectively; a radial basis function is used. A one-to-many support vector machine multi-classifier is constructed, and the penalty parameter and kernel parameter are optimized by grid search to obtain the optimal model. In the evaluation phase, the four-dimensional feature vector output from step S4 is received, normalized in the same way, and then input into the model. The model outputs the decision value for each category, and the category corresponding to the maximum value is taken as the wear level.

[0048] In summary, this algorithm for visual recognition and wear assessment of the tooth edges of a grinding machine wheel solves the problems of low accuracy in tooth edge recognition and inaccurate wear assessment by employing sub-pixel edge localization, outlier repair, and multi-dimensional feature fusion. Sub-pixel interpolation technology overcomes the physical limitations of hardware resolution, improving measurement accuracy. Addressing the characteristics of easily broken and blurred wear edges, a confidence level division and RANSAC repair mechanism are designed to effectively distinguish normal contours from wear defects and avoid noise interference. Abandoning a single judgment criterion, a multi-dimensional feature vector including deviation, curvature, area, and roughness is constructed, and a machine learning model is used for comprehensive evaluation, resulting in more objective and accurate results.

[0049] The above description is only a specific embodiment of the present invention, but the structural features of the present invention are not limited thereto. Any changes or modifications made by those skilled in the art within the scope of the present invention are covered by the patent scope of the present invention.

Claims

1. An algorithm for visual recognition and wear assessment of the tooth edges of a grinding wheel in a grinding machine, characterized in that... The following steps are included: S1: Image acquisition and region of interest extraction. The original image of the grinding wheel teeth of the grinding machine is acquired by an industrial camera under preset light conditions. Based on the geometric arrangement features of the grinding wheel teeth and a preset template matching algorithm, the region of interest image containing one or more grinding wheel teeth is accurately extracted from the original image and used as the input image for subsequent processing. S2: Image preprocessing and enhancement. The image of the region of interest is converted to grayscale, and a multi-scale noise reduction algorithm based on wavelet transform is used to remove noise during the image acquisition process. Then, an adaptive contrast enhancement method based on histogram equalization is applied to highlight the grayscale difference between the edge of the grinding wheel teeth and the background, and the enhanced grinding wheel teeth image is obtained. S3: Robust localization and anomaly identification of grinding wheel tooth edges. After receiving the enhanced grinding wheel tooth image, the Canny edge detection operator is first used to perform pixel-level coarse edge localization to obtain an initial edge point set. Then, for each edge point in the initial edge point set, the consistency of its local gradient direction is calculated to generate an edge confidence map. Based on the preset confidence threshold, the initial edge point set is divided into a smooth region point set and a high-risk anomaly region point set. For a smooth region point set, cubic spline interpolation is performed on the gray values ​​along the gradient direction to reconstruct the gray value distribution surface in the neighborhood of the point. The sub-pixel coordinates of the edge are accurately located by solving the points where the second derivative of the surface is zero, thus obtaining the first sub-pixel contour point set. For the high-risk anomaly point set, a robust fitting method based on the random sampling consensus algorithm is adopted. Through iterative random sampling and model verification, a low-order spline curve representing the macroscopic geometric trend of the grinding wheel teeth is fitted. Using this curve as a benchmark, points that deviate from the curve by more than a preset distance threshold are marked as anomaly point sets. At the same time, the part of the curve corresponding to the high-risk anomaly area is sub-pixel encrypted to obtain the second sub-pixel contour point set. Finally, the first sub-pixel contour point set and the second sub-pixel contour point set are merged to generate the repaired sub-pixel level macroscopic contour point set, and the abnormal point set is output separately. S4: Extraction of wear feature parameters of grinding wheel teeth. The system receives the repaired sub-pixel level macroscopic contour point set and the abnormal point set. Based on the repaired sub-pixel level macroscopic contour point set, it fits the standard geometric contour of the grinding wheel teeth using the least squares method and calculates the deviation between the actual contour and the standard contour. Simultaneously, it extracts the local curvature change and wear area of ​​the grinding wheel tooth edge based on the abnormal point set, and calculates the edge roughness based on the repaired sub-pixel level macroscopic contour point set. The deviation, local curvature change, wear area, and edge roughness are used together as feature parameters to quantify the degree of wear. S5: Comprehensive assessment and grading of wear degree. It receives the deviation, local curvature change, wear area and edge roughness extracted in step S4 and inputs them into a pre-built support vector machine-based wear degree assessment model. This model is trained with historical wear data and outputs the wear degree level of the grinding wheel teeth. The wear degree level includes slight wear, moderate wear and severe wear.

2. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: In step S1, image acquisition and region of interest extraction use blue monochromatic light with a wavelength of 450nm to 460nm as a preset light source to eliminate ambient light interference and enhance the contrast of the grinding wheel tooth edges. The industrial camera has a resolution of more than 5 million pixels and acquires the original image when the grinding wheel is stationary. The template matching algorithm calculates the similarity between the original image and the pre-stored standard tooth template based on the normalized cross-correlation function. The region corresponding to the peak similarity position is the region of interest. It is cropped from the original image, rotated to the standard pose, and then output to step S2.

3. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: The wavelet transform-based multi-scale denoising algorithm in step S2 converts the region of interest image into a grayscale image, and then performs three-level two-dimensional discrete wavelet decomposition using the Daubechies-4 wavelet basis to obtain the low-frequency approximate components. And high-frequency detail components in the horizontal, vertical and diagonal directions. ,in Indicates the number of decomposition layers; applies a soft thresholding function to shrink the coefficients of the high-frequency detail components in each direction of each layer. ;in These are the original wavelet coefficients. The threshold adjustment factor is set to 2.

5. The noise standard deviation estimate is calculated using the median absolute deviation method; finally, the denoised image is reconstructed through wavelet inverse transform and output to the subsequent histogram equalization process.

4. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: The adaptive contrast enhancement method in step S2 specifically involves receiving the wavelet-denoised image, dividing it into several non-overlapping local sub-blocks, each sub-block being 64×64 pixels in size; calculating the grayscale histogram for each sub-block separately, and setting a cropping limiting coefficient. The gray levels in the histogram that are higher than the total number of pixels in the sub-blocks are cropped and the cropped parts are evenly redistributed to all gray levels to limit contrast amplification. Then, histogram equalization is performed on each sub-block, and bilinear interpolation is used to eliminate the block effect at the sub-block boundary. Finally, the contrast-enhanced grinding wheel tooth image is output to step S3.

5. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: The specific process of generating the edge confidence map and dividing the point set in step S3 is as follows: receiving the enhanced image output from step S2, firstly, the Canny operator is used to extract pixel-level edge points to obtain the initial edge point set. For each edge point Take a 5×5 neighborhood window centered on it, and calculate the gradient direction angle of all pixels within the window. ,in Given the pixel coordinates within the neighborhood; then calculate the variance of the gradient direction angle. in The mean of the gradient direction angles within the neighborhood; the variance The reciprocal of the normalized value is used as the confidence level for that point. Generate an edge confidence map; preset confidence threshold. ,Will The points are assigned to the smooth region point set. ,Will The points are classified into the high-risk anomaly point set. The results are then output to subsequent processing.

6. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 5, characterized in that: In step S3, the smooth region point set is received. For each of the edge points Cubic spline interpolation is performed along the gradient direction with a step size of 0.15 pixels to reconstruct the gray-level distribution function along the normal direction of the point. ,in Let the coordinates be along the normal direction; solve the equation. Obtain the precise location of the grayscale inflection point Mapping it back to the image coordinate system yields subpixel coordinates. For all Perform the above operations on the points in the matrix to generate the first sub-pixel contour point set. The output is then fed into subsequent fusion steps to complete the sub-pixel localization of the point set in the smooth region.

7. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 5, characterized in that: In step S3, the robust fitting process for the high-risk anomaly region point set involves receiving the high-risk anomaly region point set. A robust fitting method based on the random sampling consensus algorithm is adopted, and a maximum number of iterations is set. Interior point distance threshold Pixels; in each iteration, from Three points are randomly selected, and a second-order spline curve is fitted as a candidate model. ;calculate All points The Euclidean distance, the statistical distance is less than The points are taken as interior points, and the number of interior points is recorded; After the iteration is complete, the candidate model with the most interior points is selected as the optimal model. That is, a low-order spline curve characterizing the macroscopic geometric trend of the grinding wheel teeth; medium to The distance is greater than The points are marked as anomaly point set At the same time exist The corresponding intervals are sampled with encrypted sampling at 0.1 pixel intervals to obtain the second sub-pixel contour point set. ; final fusion and Generate the repaired subpixel-level macroscopic contour point set and will Output separately to step S4.

8. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: The wear feature parameter extraction in step S4 specifically includes: receiving the repaired sub-pixel level macroscopic contour point set. and anomaly set ;based on The standard geometric profile is fitted using the least squares method. Let the equation of the standard profile be... ,for Each point in Calculate its directed distance to the standard profile: ; Recalculate the deviation characteristics: ； Local curvature variation characteristics Through calculation The standard deviation of the curvature of the arc formed by every three adjacent points is obtained; the area characteristics of the wear zone are also described. Through calculation The total pixel area of ​​the connected component formed by all outliers is obtained by specifically... Perform convex hull detection and calculate the convex hull area; edge roughness features. based on Calculation, in For each point in the standard profile, take a window of length 11 points along the edge normal. Calculate the root mean square (RMS) distance from all points within the window to the standard profile, and then average the RMS distances of all points to obtain the final profile. ;Will As a four-dimensional feature vector, it is output to step S5.

9. The algorithm for visual recognition and wear assessment of the tooth edge of a grinding wheel as described in claim 1, characterized in that: The support vector machine model construction and training in step S5 specifically includes: pre-collecting 500 historical grinding wheel tooth samples each of three categories: slight wear, moderate wear, and severe wear; and extracting the feature vector of each sample through steps S1 to S4. The training dataset is constructed; Z-score normalization is performed on each feature dimension, i.e. ,in and These are the mean and standard deviation of the feature on the training set, respectively; a radial basis function is used. Construct a one-to-many support vector machine multi-classifier and optimize the penalty parameter using a grid search method. and kernel parameters Finally, the optimal model is obtained; in the evaluation phase, the four-dimensional feature vector output from step S4 is received, and after the same normalization, it is input into the model. The model outputs the decision value for each category, and the category corresponding to the maximum value is taken as the wear level.

Images