Intelligent method and system for visual-based bearing roller surface and profile detection

Abstract

The application discloses a kind of based on visual bearing roller surface and profile Intelligent detection method and system, method includes constructing world coordinate system, based on the image acquisition device of different station Synchronous acquisition bearing roller in the same position Surface image and two-dimensional profile image;Respectively to surface point cloud and profile edge point cloud are fitted and are carried out rough alignment and are verified and are finely registered, obtain registration surface point cloud and registration profile edge point cloud;Based on correction surface point cloud and cylindrical surface model Generation to be measured surface development drawing And corresponding coordinate mapping relationship, mathematical modeling and calculus analysis are carried out to correction profile edge point cloud, obtain continuous smooth bearing roller Actual profile curve, obtain profile analysis result;Based on surface defect result and profile analysis result, obtain determination result.The application can be realized in non-contact condition Synchronous realization bearing roller surface micro defect high-precision detection and logarithm parent line profile micrometer level detection and submicron level profile evaluation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a vision-based intelligent detection method and system for the surface and contour of bearing rollers. Background Technology

[0002] As the most fundamental supporting component in rotating machinery, the performance of bearings largely depends on the geometric accuracy and surface condition of the rolling elements themselves. In current technology, high-end bearing designs mostly employ logarithmic generatrix shaping technology, which precisely shapes the bearing roller generatrix according to a logarithmic curve to achieve a more uniform load distribution along the axial direction and avoid stress concentration at both ends. However, achieving this effect requires controlling the contour accuracy to the sub-micron level and being able to detect minute surface defects such as microcracks, small pits, or grinding burns. These defects are often only a few micrometers to sub-micrometers in size.

[0003] Currently, two main inspection technologies are used: traditional contact profilometers or mechanical gauges, and non-contact optical inspection technology. Contact profilometers or mechanical gauges have several problems: they easily scratch the surface of bearing rollers; their inspection efficiency is too low, making it difficult to achieve 100% comprehensive inspection on the production line; and they are difficult to simultaneously assess both shape accuracy and surface quality.

[0004] While non-contact optical inspection technology can solve the problem of contact damage, simultaneously performing surface defect detection and sub-micron level assessment of logarithmic generatrix profiles within the same system still faces numerous practical challenges. The main contradiction lies in the fact that defect detection requires high lateral resolution and sufficient depth of field to capture three-dimensional morphology; while assessing generatrix profiles demands nanometer-level longitudinal resolution over measurement lengths of tens to hundreds of millimeters, and requires the construction of accurate logarithmic curve models to calculate profile deviations. Existing non-contact solutions often struggle to balance both, either sacrificing resolution for a large measurement range or limiting measurement to a small area to ensure high accuracy.

[0005] In other words, there is currently no mature technology to unify defect information and contour data under the same coordinate system, making it impossible to acquire and correlate the two simultaneously. The dilemma of not being able to consider the contour while detecting defects, and missing defects while detecting the contour, severely restricts the closed-loop quality control of bearing rollers and has become a bottleneck problem that urgently needs to be solved in the intelligent manufacturing of high-end bearings. Although existing technologies can perform surface defect detection or contour detection separately, the lack of a unified spatial representation method for different detection data makes it difficult to establish a spatial correspondence between surface defects and contour errors, resulting in the inability to reliably correlate and collaboratively determine the detection results. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing a vision-based intelligent detection method and system for the surface and contour of bearing rollers. This invention does not simply combine surface and contour detection; instead, it establishes a unified spatial coordinate system to achieve unified expression, spatial correlation, and collaborative analysis of multi-source detection data, thereby completing a comprehensive quality evaluation of the bearing rollers.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A vision-based intelligent detection method for bearing roller surfaces and contours includes the following steps: A world coordinate system is constructed, and surface images and two-dimensional contour images of the bearing rollers at the same location are acquired simultaneously. These images are then preprocessed to obtain surface point clouds and contour edge point clouds in the world coordinate system. The surface point cloud and the contour edge point cloud are fitted, aligned and registered respectively to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. The surface unfolding diagram and corresponding coordinate mapping relationship of the surface to be tested are generated based on the registered surface point cloud, and the surface detection process and the contour detection process use the same spatial coordinate reference. An enhanced memory library is constructed based on a defect-free surface unfolding atlas. The characterization features of the unfolding map of the surface to be tested are extracted, the surface defect region is determined, the corresponding spatial position is recovered through coordinate mapping relationship, and the defect features are calculated to obtain the surface defect result. Based on the registration contour edge point cloud, the actual contour curve of the bearing roller is reconstructed, the contour deviation between the actual contour curve and the theoretical contour curve is calculated, the contour abnormal area and the corresponding world coordinate are determined, and the contour analysis result is formed. Based on the coordinate mapping relationship, a spatial correlation relationship is constructed between the surface defect region and the contour anomaly region. The spatial correlation relationship is used to determine whether the surface defect region and the contour anomaly region correspond to the same physical region. Cross-validation and fusion analysis are performed only on the surface defect results and contour analysis results that correspond to the same physical region. The comprehensive quality judgment is completed through a multi-level decision model.

[0008] As one possible implementation method, the surface point cloud and contour edge point cloud are obtained through the following steps: A world coordinate system is constructed by taking the ideal axial direction of the bearing rollers as the Z-axis and the radial plane as the XY plane. In response to the triggering of the bearing roller reaching the preset detection position, the surface imaging station and the contour imaging station are synchronously controlled to acquire surface images and two-dimensional contour images respectively. Select the optimal surface image, convert the coordinates of each pixel in the optimal surface image into the initial distortion-free normalized coordinates in the world coordinate system, and perform correction processing to obtain the corrected image coordinates; Based on the corrected image coordinates, the corresponding ray direction, and the prior model of the cylinder, the intersection point of the ray and the cylinder surface is calculated, and the intersection point closest to the camera is selected as the corresponding surface point to obtain the three-dimensional coordinates of the surface point in the world coordinate system, thus forming a surface point cloud. A two-dimensional contour image is acquired, sub-pixel edges are extracted and distortion correction is performed to obtain distortion-free coordinates. The distortion-free coordinates are then transformed to the world coordinate system using camera intrinsics to obtain the three-dimensional coordinates of the contour edge points, thus forming a contour edge point cloud.

[0009] As one possible implementation, the process of fitting, aligning, and registering the surface point cloud and the contour edge point cloud to obtain the registered surface point cloud and the registered contour edge point cloud includes the following steps: Cylindrical fitting is performed on the surface point cloud and the contour edge point cloud respectively to obtain the radius and axis of the corresponding cylinder. It is then determined whether the fitted radius and axis are within the preset tolerance range. If they are within the preset tolerance range, coarse alignment is completed. The optimal rigid body transformation is determined by the least squares method. Based on the optimal rigid body transformation, the surface point cloud and the contour edge point cloud are finely registered to unify the surface point cloud and the contour edge point cloud to the same spatial coordinate reference, thus obtaining the registered surface point cloud and the registered contour edge point cloud. Maintain the spatial correspondence between the registered surface point cloud and the registered contour edge point cloud so that surface texture information, contour geometry information and subsequent detection results are all expressed based on a unified spatial coordinate system; Each surface point cloud includes corresponding three-dimensional coordinates and associated original texture color. The contour edge point cloud is used to characterize the actual geometric contour of the bearing roller. The optimal rigid body transformation is the rigid body transformation that minimizes the overall distance between the surface point cloud and the contour edge point cloud in the common geometric feature region.

[0010] As one possible implementation, the process of generating the unfolded map of the surface under test and the corresponding coordinate mapping relationship based on the registered surface point cloud includes the following steps: Cylindrical fitting is performed on the registered surface point cloud, and the cylinder parameters are optimized by nonlinear least squares method to minimize the sum of squares of the deviations between the distances from each 3D point on the surface to the axis of the fitted cylinder and the fitted radius, thus obtaining the optimal cylinder parameters. The cylinder parameters include the axial direction, a point on the axis, and the radius. Based on the three-dimensional points of each surface and the optimal cylinder parameters, the cylinder coordinates corresponding to the three-dimensional points of each surface are calculated. The cylinder coordinates include axial coordinates and circumferential angle coordinates. The vertical pixel position is determined according to the axial coordinates and the preset axial pixel density. The horizontal pixel position is determined according to the circumferential angle coordinates and the total pixel width of the cylinder unfolded diagram. The circumferential boundary is then processed by wrapping around it. The size of the cylindrical unfolded pattern is determined based on the axial coordinate range and axial pixel density. The cylindrical unfolded pattern is generated by combining the color information corresponding to each horizontal and vertical pixel position. The missing areas are filled by interpolation based on the color information of neighboring pixels to obtain a continuous and complete surface unfolded pattern. A one-to-one correspondence is established between the pixels of the cylindrical unfolded image, the three-dimensional coordinates, and the cylindrical coordinates to form a coordinate mapping relationship. The coordinate mapping relationship is then used to achieve a unified positional mapping between the unfolded image pixels, the surface point cloud, and the contour edge point cloud, providing a unified spatial positioning benchmark for determining whether the surface defect area and the contour abnormal area correspond to the same physical area.

[0011] As one possible implementation, the construction of the enhanced memory library based on the defect-free surface unfolding atlas includes the following steps: Feature extraction is performed on the unfolded atlas of defect-free surfaces to obtain at least two feature maps of different scales, and multidimensional feature vectors corresponding to the spatial locations of each feature map are extracted respectively; Local feature aggregation is performed on each feature vector, the statistical features of the feature vectors in the preset neighborhood are calculated, and normalization is performed to obtain normalized feature vectors. The normalized feature vectors are combined to form the feature vector set of the corresponding surface unfolded map. The feature vector sets corresponding to each surface unfolding map are collected to form a global feature pool. Cosine distance is used as the distance metric. A preset number of representative features are selected from the global feature pool according to a preset selection criterion to build an initial memory bank. The features of normal samples collected under different working conditions are added to the initial memory, and the initial memory is enhanced to obtain the enhanced memory.

[0012] As one possible implementation, the step of extracting the characterization features of the unfolded image of the surface to be tested, determining the surface defect region, recovering the corresponding spatial position through coordinate mapping relationship, and calculating the defect features to obtain the surface defect result includes the following steps: The characterization features of the surface unfolded map to be tested are extracted and similarity calculation is performed with the augmented memory. The cosine distance is used to calculate the nearest neighbor distance between each feature to be tested and the corresponding feature in the augmented memory. The nearest neighbor distance is used as the anomaly score of the corresponding region. Each anomaly score is mapped to the spatial location of the corresponding feature map to obtain anomaly score maps at different scales; each anomaly score map is upsampled to the same size as the original unfolded map, and then fused by taking the maximum value at the pixel level to generate an anomaly heatmap; The Otsu threshold is calculated based on the abnormal heat map, and the segmentation threshold is determined by multiplying the Otsu threshold by a preset magnification factor to generate a binary defect mask. The binarized defect mask is subjected to closing and opening operations in sequence to extract connected components and filter out regions with an area smaller than a preset area threshold to obtain the target defect region. By mapping coordinates, the pixel coordinates corresponding to the target defect area are restored to the world coordinate system and cylindrical coordinate system. The spatial location, geometric features and texture features of the target defect area are calculated. Based on the classifier trained by geometric features, texture features, color features, shape features and context features, the defect type is identified, and the surface defect result is formed.

[0013] As one possible implementation, constructing the spatial association between surface defect regions and contour anomaly regions based on coordinate mapping includes the following steps: Based on the coordinate mapping relationship, the pixel coordinates corresponding to the surface defect area are converted into world coordinates and cylindrical coordinates, and the contour anomaly area and its corresponding world coordinates are determined based on the contour analysis results. The axial distance, circumferential distance, and three-dimensional spatial distance between the surface defect area and the contour abnormal area are calculated respectively. The circumferential distance is calculated based on the difference between the circumferential angular coordinates of the two areas, and the minimum angular distance is taken in combination with the circumferential surrounding characteristics of the cylindrical surface. It is determined whether the axial distance is less than the preset axial tolerance, the circumferential distance is less than the preset circumferential tolerance, and the three-dimensional spatial distance is less than the preset spatial tolerance. If the axial distance, circumferential distance, and three-dimensional spatial distance all meet the corresponding tolerance requirements, it is determined that the surface defect area and the contour abnormal area belong to the same physical area, and a corresponding associated defect pair is constructed. Otherwise, no associated defect pair is constructed. Spatial consistency, positional continuity, and association confidence are calculated for each pair of associated defects. The association confidence is determined by combining spatial consistency, positional continuity, and defect type to construct spatial association relationships. It is also determined whether the surface defect area and the contour anomaly area correspond to the same physical area, which serves as the basis for cross-validation, fusion analysis, and comprehensive quality judgment.

[0014] As one possible implementation, the method includes the following steps: Based on the registered contour edge point cloud, the axis and axis direction are determined, the radial distance corresponding to each contour point is calculated, and the points are sorted according to the axial coordinates to form an ordered contour point set; outlier points are removed by the median absolute deviation filtering method to obtain the contour data after gross error removal. Based on the contour data, a discrete contour curve is constructed, and a smooth spline is used for fitting. An objective function including a fitting error term and a smoothing constraint term is constructed, and the actual contour curve is obtained by optimization. The scaling factor, axial translation, and radial translation between the actual contour curve and the theoretical contour curve are optimized using the least squares method to eliminate the rigid displacement error between the two and obtain the registered actual contour curve. The contour deviation, derivative characteristics, and curvature characteristics between the registered actual contour curve and the theoretical contour curve are calculated, and multi-level contour analysis is performed based on the contour deviation, derivative characteristics, and curvature characteristics. Based on the results of multi-level contour analysis, abnormal contour regions are identified, and the world coordinate positions corresponding to the abnormal contour regions are determined. Contour defect diagnostic information is formed by combining the continuity, change trend and distribution characteristics of the abnormal contour regions, wherein the change characteristics include change trend and distribution characteristics. By combining contour accuracy indicators and contour defect diagnosis information, contour analysis results are generated.

[0015] As one possible implementation method, the quality determination through a multi-level decision model includes the following steps: A multi-level decision-making model is constructed, which includes a surface defect judgment unit, a contour quality judgment unit, and a fusion judgment unit. The surface quality evaluation result is calculated by the surface defect determination unit based on the defect type, defect area range and defect confidence level. The contour quality evaluation result is calculated by the contour deviation, derivative characteristics, curvature characteristics and contour abnormal area by the contour quality judgment unit. The contour quality judgment adopts the protection zone principle, calculates the expanded uncertainty based on the standard uncertainty and coverage factor of the contour index, and makes a judgment based on the expanded uncertainty correction tolerance zone. Based on the spatial correlation, the fusion judgment unit performs cross-validation and fusion analysis on the surface quality evaluation results and the contour quality evaluation results. If there are correlated defect pairs and the severity combination of the correlated defect pairs exceeds the preset correlation threshold, it is directly judged as unqualified; otherwise, the final quality judgment result is calculated based on the comprehensive weighted calculation of the surface quality evaluation results and the contour quality evaluation results to obtain the final quality judgment result of the bearing roller under test.

[0016] A vision-based intelligent inspection system for bearing roller surfaces and contours, comprising: A generation module is constructed to build a world coordinate system, simultaneously acquire surface images and two-dimensional contour images of the bearing rollers at the same location, and preprocess them respectively to obtain surface point clouds and contour edge point clouds in the world coordinate system. The alignment and registration module performs fitting, alignment, and registration on the surface point cloud and the contour edge point cloud respectively, to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. The mapping generation module generates the unfolded map of the surface to be tested and the corresponding coordinate mapping relationship based on the registered surface point cloud, and keeps the surface detection process and the contour detection process using the same spatial coordinate reference. The identification module is determined by constructing an enhanced memory library based on the defect-free surface unfolding atlas, extracting the characterization features of the surface unfolding atlas to be tested, identifying the surface defect region, restoring the corresponding spatial position through coordinate mapping relationship and calculating the defect features to obtain the surface defect result; The module is determined, and the actual contour curve of the bearing roller is reconstructed based on the registered contour edge point cloud. The contour deviation between the actual contour curve and the theoretical contour curve is calculated, the contour anomaly area and its corresponding world coordinates are determined, and the contour analysis results are generated. The fusion analysis module is used to construct the spatial relationship between surface defect areas and contour anomaly areas based on coordinate mapping. The spatial relationship is used to determine whether surface defect areas and contour anomaly areas correspond to the same physical area. Cross-validation and fusion analysis are performed only on surface defect results and contour analysis results that correspond to the same physical area. The comprehensive quality judgment is completed through a multi-level decision model.

[0017] This invention, by adopting the above technical solutions, has significant technical effects: This invention establishes a unified spatial coordinate system, realizing the unified expression and spatial mapping of surface and contour information acquired from different inspection stations. This enables the establishment of a physical correspondence between surface defects and contour errors, and on this basis, completes spatial correlation and fusion analysis, thereby achieving the coordinated detection of bearing roller surface quality and contour accuracy, and improving the accuracy, reliability and comprehensive evaluation capability of the detection results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figures 2-9 This is a technical effect diagram of a specific embodiment of the present invention; Figure 10 This is a schematic diagram of the overall structure of the system of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

[0021] Example 1: A vision-based intelligent detection method for bearing roller surfaces and contours, such as... Figure 1 As shown, it includes the following steps: S100. Construct a world coordinate system, simultaneously acquire surface images and two-dimensional contour images of the bearing rollers at the same location, and perform preprocessing respectively to obtain surface point clouds and contour edge point clouds in the world coordinate system. S200. Fit, align and register the surface point cloud and the contour edge point cloud respectively to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. S300: Generate the unfolded map of the surface to be tested and the corresponding coordinate mapping relationship based on the registered surface point cloud, and keep the surface detection process and the contour detection process using the same spatial coordinate reference. S400: Construct an enhanced memory library based on the defect-free surface unfolding atlas, extract the characterization features of the surface unfolding atlas to be tested, determine the surface defect region, recover the corresponding spatial position through coordinate mapping relationship and calculate the defect features to obtain the surface defect result; S500: Based on the registered contour edge point cloud, reconstruct the actual contour curve of the bearing roller, calculate the contour deviation between the actual contour curve and the theoretical contour curve, determine the contour abnormal area and the corresponding world coordinates, and form the contour analysis result. S600. Based on the coordinate mapping relationship, construct the spatial association between the surface defect area and the contour abnormal area. The spatial association is used to determine whether the surface defect area and the contour abnormal area correspond to the same physical area. Cross-validation and fusion analysis are performed only on the surface defect results and contour analysis results that correspond to the same physical area. The comprehensive quality judgment is completed through a multi-level decision model.

[0022] In one embodiment, constructing a world coordinate system, synchronously acquiring surface images and two-dimensional contour images of the same location on the bearing roller in response to a trigger signal, and performing preprocessing on each to obtain surface point clouds and contour edge point clouds in the world coordinate system includes the following steps: A world coordinate system is constructed by taking the ideal axial direction of the bearing rollers as the Z-axis and the radial plane as the XY plane. In response to the triggering of the bearing roller reaching the preset detection position, the surface imaging station and the contour imaging station are synchronously controlled to acquire surface images and two-dimensional contour images respectively. Select the optimal surface image, convert the coordinates of each pixel in the optimal surface image into the initial distortion-free normalized coordinates in the world coordinate system, and perform correction processing to obtain the corrected image coordinates; Based on the corrected image coordinates, the corresponding ray direction, and the prior model of the cylinder, the intersection point of the ray and the cylinder surface is calculated, and the intersection point closest to the camera is selected as the corresponding surface point to obtain the three-dimensional coordinates of the surface point in the world coordinate system, thus forming a surface point cloud. A two-dimensional contour image is acquired, sub-pixel edges are extracted and distortion correction is performed to obtain distortion-free coordinates. The distortion-free coordinates are then transformed to the world coordinate system using camera intrinsics to obtain the three-dimensional coordinates of the contour edge points, thus forming a contour edge point cloud.

[0023] Step S100 can be understood as defining the world coordinate system and how to implement the synchronization triggering mechanism. In a specific embodiment, a unified spatiotemporal reference system is first constructed. The role of this spatiotemporal reference system is to ensure that the data collected by the two workstations correspond to the same physical location and the same time of the workpiece. The specific process of achieving time synchronization is as follows: From a hardware perspective, the entire system is coordinated by a main programmable logic controller (PLC). When the sensors used in conjunction, such as photoelectric sensors, detect that the bearing rollers on the conveyor belt have reached the preset position, the PLC synchronously sends hardware trigger signals (such as TTL pulses) to the image acquisition devices and light sources at the two imaging stations. The two imaging stations are the surface imaging station and the contour imaging station, respectively. The image acquisition device is a camera, so there are at least two cameras, one to acquire surface images and the other to acquire two-dimensional contour images.

[0024] The trigger time can be set to The exposure start times for the two cameras are respectively and Then it needs to satisfy: in, The inherent trigger jitter coefficient of the system is typically required. It is much smaller than the distance the bearing rollers travel on the conveyor belt (e.g., in At a given speed, the distance moved within 1 ms is 0.5 mm, so a camera capturing a photo can be considered an "instantaneous" capture.

[0025] To achieve the same physical location, a spatial reference needs to be preset, specifically: Construct a world coordinate system This world coordinate system Defined as a Cartesian coordinate system fixed to the testing platform, the ideal axial direction of the bearing rollers is set as... The axis and radial plane are set as Plane. The actual position of the bearing rollers is used as the workpiece coordinates, i.e., in... At that moment, the actual position and orientation of the bearing rollers are determined, and all subsequent measurements are based on the bearing rollers' relationship with the world coordinate system at that instant. The relative relationships are discussed.

[0026] The image acquisition device based on the surface imaging station captures high dynamic range texture information. The surface imaging station aims to obtain the cylindrical surface development image of bearing rollers to detect surface defects such as oil stains and scratches. The mathematical model used is the pinhole imaging model and the multispectral imaging principle.

[0027] This workstation uses an area scan camera, also known as a surface camera in other embodiments, following a central perspective projection model, assuming an object point in the world coordinate system. The coordinates in are In the camera coordinate system, the coordinates are (in, (This is an extrinsic parameter), and the projection of this object point onto the normalized image plane is: .

[0028] After considering radial and tangential distortions, the distortion-free normalized coordinates are obtained. Final pixel coordinates Represented as: in, This represents the intrinsic parameter matrix of the surface camera corresponding to the image acquisition device at the surface imaging station. This intrinsic parameter matrix is ​​obtained through calibration.

[0029] Because oil stains exhibit different characteristics under different lighting conditions (such as specular highlights and interference colors), multi-angle or multi-band illumination is used to enhance the contrast between oil stain defects and normal areas. This ensures that oil stain defects of varying depths and shapes are clearly displayed, avoiding missed detections and misjudgments. The specific process involves switching light source modes and sequentially activating illumination from different angles or wavelengths to obtain images of the same area under different physical lighting conditions. , Indicates wavelength (e.g., RGB). This represents the angle of incidence of the light. The entire process can acquire N images. The image with the best contrast can be selected directly, or an enhanced image can be generated through pixel-level fusion (such as taking the maximum value or weighted averaging). : in, The weights are obtained by optimizing according to different defect types. Represents the coordinates of each pixel. Indicates the first Image.

[0030] Ultimately, one or more preliminary surface images can be obtained. Where C is the number of channels (usually 3, i.e., RGB), and each pixel in the surface image corresponds to the world coordinate system. Every tiny area on the surface of the bearing rollers.

[0031] The image acquisition device based on the contour imaging station captures the high-precision edge position of the bearing roller. This station aims to obtain the precise geometric contour of the bearing roller's cross-section and uses a line scan camera (also referred to as a contour camera in other embodiments). The bearing roller is placed between a high-uniformity backlight and the line scan camera. The entire process relies on backlight projection and line scan scanning. Specifically, because the bearing roller is placed between the high-uniformity backlight and the line scan camera, ideally, the part of the bearing roller that is obscured forms a dark area, while the background is a bright area, producing a sharply changing edge on the image.

[0032] A line scan camera captures one line of pixels per exposure, so a rotating mechanism can be used to rotate the bearing rollers around its axis at a constant angular velocity. Rotation; or relative linear motion between the line scan camera and the bearing rollers, with a time difference of... Then obtain the time The inner scan yielded a length of A sequence of linear array images is used to stitch together a two-dimensional contour image. .

[0033] This embodiment extracts sub-pixels from a 2D contour image using a sub-pixel edge preliminary localization model. For a detectable edge in the 2D contour image, a single row of pixels is extracted along a direction perpendicular to this edge; this is the single row of scan lines. Arranging the gray values ​​of each pixel position within this single row of scan lines in spatial order forms a one-dimensional grayscale profile. Ideally, the grayscale profile at the edge exhibits a step-like change, with the midpoint of this step representing the ideal edge location. However, due to optical blurring and discrete sampling, the actual grayscale transition is a continuous, gradual change rather than an ideal step. Therefore, the sub-pixel coordinates of this step midpoint can be estimated by fitting an error function model or calculating the grayscale moments.

[0034] Specifically, within the pixel interval spanning the edge, using pixel coordinates as positional weights and the grayscale value after deducting the dark background as the quality, the centroid coordinates of this quality distribution are calculated. The obtained centroid coordinates are the sub-pixel positions of the edge. , is represented as: ,in, It is a pixel range that spans the edge. It is the average gray level of the dark area.

[0035] For two-dimensional contour images For each column (corresponding to an angle on the circumference of the bearing roller), the above sub-pixel edge detection is performed to obtain a pair of left and right edge points. ,in, It is the scanning direction coordinate (corresponding to the axial position of the bearing roller).

[0036] Camera intrinsic parameter matrix using contour imaging station And the known backlight plane in the world coordinate system Location , each pixel coordinate Back projection onto the world coordinate system For the left edge point, we have: in, Indicates the scale factor. This represents the camera extrinsic parameters of the linear scan camera at the contour imaging station. For each axial position... A pair of world coordinate points is calculated, which define the diameter direction of the cross section. By combining the points of all cross sections, the contour edge point cloud is obtained. .

[0037] Finally, the spatiotemporal alignment is verified to determine whether the basic dimensions of the bearing rollers calculated by the two workstations (such as length and maximum radius) are within the tolerance range, in order to confirm the accuracy of the trigger synchronization and calibration.

[0038] Finally, step S100 yields a data packet containing: a surface image. and their corresponding camera parameters Outline edge dot cloud and its corresponding scanning parameters (such as rotation angle and axial position mapping relationship), trigger time and workpiece ID.

[0039] Step S100 converts the physical properties (surface texture, geometric edges) of the bearing rollers into digital information with precise spatiotemporal labels and physical coordinate mapping relationships, providing a reliable and traceable data foundation for subsequent steps such as coordinate unification, AI analysis, and calculus calculation.

[0040] The next step, S200, involves intelligent detection and localization of surface defects on the cylindrical surface of the bearing rollers. The core objective is to identify and locate various surface anomalies, including oil stains, scratches, and pitting, from the pre-processed surface development image. The specific process is as follows: First, the cylindrical surface model is precisely fitted and the surface unfolded map is generated. Based on the contour edge point cloud obtained in step S100, the actual cylindrical surface parameters of the bearing roller are fitted. The theoretical shape of the bearing roller is a standard cylinder, and its parameters can be defined as: axial direction vector, coordinates of any point on the axis, and cylinder radius R. Using the minimum distance deviation from all points in the contour edge point cloud to the fitted cylindrical surface as the optimization objective, the optimal fitting parameters are obtained through least squares iteration, completing the accurate reconstruction of the actual cylindrical surface and avoiding coordinate mapping deviations caused by machining errors and clamping offsets. After fitting, a cylindrical coordinate system is constructed with the fitted cylinder axis as the Z-axis. Any point on the cylindrical surface can be uniquely represented by two parameters: axial coordinate z and circumferential angle θ. All texture pixels on the three-dimensional cylindrical surface are unfolded and arranged sequentially in the order of (z, θ) onto a two-dimensional plane, thus obtaining a cylindrical surface unfolded map without geometric distortion. This ensures that the texture features of the same area are continuous and complete during subsequent defect detection, without splicing breaks or dimensional deformation.

[0041] After generating the unfolded image, an improved PatchCore algorithm is used for anomaly detection. First, during the training phase, the model is trained using only unfolded images of normal bearing rollers without defects. Texture feature memory is extracted from normal samples, allowing model construction without relying on a large number of labeled defective samples, thus adapting to the limited defect samples characteristic of industrial scenarios. In the inference phase, the unfolded image of the bearing roller to be detected is input into the trained model. Features are extracted from each local image patch and compared with normal features in the memory. Anomaly scores are calculated for each location, generating an anomaly score heatmap of the same size as the original image. Areas with scores higher than a preset threshold are identified as defective regions, achieving accurate anomaly region localization. To address the specificity of oil stain defects, an additional classification step is added: based on the contrast difference between oil stains and real physical defects in different bands of the multispectral image, spectral features of the defective region are extracted. Combined with anomaly score distribution features, a lightweight classifier is trained to distinguish oil stain defects from irreversible physical defects such as scratches and pitting, avoiding misclassification of washable oil stains as defective products and improving detection accuracy.

[0042] After the above steps are completed, the surface defect detection results are output, including the location, size, type of all defect areas and the corresponding average anomaly score, providing a basis for the subsequent comprehensive quality assessment of the surface dimension.

[0043] After surface defect detection, the process proceeds to step S300, where parametric geometric accuracy analysis is performed on the bearing roller profile. The core of this analysis is based on the acquired profile edge point cloud, analyzing the deviation between the roller profile and the ideal design profile to assess whether the geometric accuracy meets the requirements. First, the acquired profile edge point cloud is preprocessed by projecting the 3D profile edge point cloud onto a cross-section perpendicular to the axis. Profile points are extracted from each cross-section, and the actual profile curve for each cross-section is fitted, completing the conversion from 3D point cloud to parametric profile curve. Next, the actual profile is registered with the theoretical profile model provided by the design, eliminating translational and rotational deviations caused by clamping, ensuring the coordinate systems of the two profiles are aligned, and providing a benchmark for subsequent deviation calculations.

[0044] After registration, a multi-order calculus core analysis is performed: First, zero-order deviation analysis is conducted to calculate the normal deviation from the theoretical contour at each angular position, obtaining the deviation distribution of the entire contour. Maximum and average deviations are statistically analyzed to determine if the contour dimensions exceed the allowable tolerance range. Next, first-order derivative analysis is performed to calculate the change of the tangent angle of the contour curve with the contour position, analyze the continuity of the contour slope, and identify abrupt changes such as local protrusions and depressions on the contour. Finally, second-order derivative and curvature analysis is performed, which is the core of contour accuracy evaluation. By calculating the curvature at each position of the contour, the curvature distribution is analyzed to determine if it meets design requirements. The rate of change of curvature, i.e., the derivative of curvature, is further calculated to evaluate the smoothness of the contour curvature and identify minor contour processing distortions.

[0045] After completing the above analysis, the uncertainty of all deviation and curvature calculation results is evaluated. Combining multiple error sources such as camera calibration error, sub-pixel edge extraction error, and registration error, the combined uncertainty of the final contour parameters is calculated, providing a reliable basis for subsequent judgment. The final output contour detection result includes parameters such as contour size deviation, curvature distribution, and uncertainty, providing a basis for subsequent comprehensive quality judgment of the contour dimensions.

[0046] Finally, proceed to step S400 to complete the comprehensive quality assessment of the bearing rollers. First, individual assessments are performed based on the surface defect detection results and the contour detection results: according to the preset surface quality standards, combined with the type, quantity, and size of defects, it is determined whether the surface quality of the bearing rollers is qualified; then, according to the requirements of contour deviation and curvature accuracy, it is determined whether the contour quality is qualified; finally, combining the two individual assessment results and the confidence level of the detection results, the final comprehensive assessment result is given, and the corresponding assessment confidence level is output, completing the entire intelligent inspection process of the bearing rollers.

[0047] In step S200, the process of fitting, aligning, and registering the surface point cloud and the contour edge point cloud to obtain the registered surface point cloud and the registered contour edge point cloud includes the following steps: Cylindrical fitting is performed on the surface point cloud and the contour edge point cloud respectively to obtain the radius and axis of the corresponding cylinder. It is then determined whether the fitted radius and axis are within the preset tolerance range. If they are within the preset tolerance range, coarse alignment is completed. The optimal rigid body transformation is determined by the least squares method. Based on the optimal rigid body transformation, the surface point cloud and the contour edge point cloud are finely registered to unify the surface point cloud and the contour edge point cloud to the same spatial coordinate reference, thus obtaining the registered surface point cloud and the registered contour edge point cloud. Maintain the spatial correspondence between the registered surface point cloud and the registered contour edge point cloud so that surface texture information, contour geometry information and subsequent detection results are all expressed based on a unified spatial coordinate system; Each surface point cloud includes corresponding three-dimensional coordinates and associated original texture color. The contour edge point cloud is used to characterize the actual geometric contour of the bearing roller. The optimal rigid body transformation is the rigid body transformation that minimizes the overall distance between the surface point cloud and the contour edge point cloud in the common geometric feature region.

[0048] In step S100, a surface image is obtained. and the corresponding camera extrinsic matrix and surface camera intrinsic parameter matrix Contour geometric data refers to two-dimensional contour images or preliminary extracted point sequences. , and contour camera intrinsic parameter matrix .

[0049] The goal of step S200 is to obtain: the three-dimensional coordinates of the surface data: for the surface image Each pixel Solve for the corresponding 3D object point in the world coordinate system. World coordinates , This represents the coordinates of a point in the world coordinate system. These represent pixels in the surface image. Represents the coordinates in the surface image; The three-dimensional coordinates of the contour data are the point series Transform points in the coordinate system to the world coordinate system In the middle, a refined outline of point clouds is formed. .

[0050] After acquiring this data, it is necessary to unify the data, that is, to share the same world coordinate system between the two sources of 3D point clouds. This is followed by fusion analysis (e.g., locating surface defects to specific axial and circumferential positions of the profile). The entire step S200 can be understood as a complete transformation from pixel to world coordinates. First, lens distortion correction is required. Since camera lenses cause radial and tangential distortion, bending straight lines, this must be corrected. The correction steps are as follows: Assume the distortion-free normalized coordinates (i.e., camera coordinates) of a 3D point in its camera coordinate system are: , , This represents the coordinates of a 3D point in the camera coordinate system. The distortion model typically uses the Brown Conrady model, represented as: in, , They represent the radial distortion coefficients, These represent the tangential distortion coefficients, which were obtained during camera calibration. This represents the radial distance from the normalized point to the optical center. The subscript indicates an even power of the radial distance. This represents the tangential distortion coefficient.

[0051] For camera images captured by the camera Each pixel coordinate We need to find the corresponding distortion-free normalized coordinates and construct the initial assumption of distortion-free coordinates, expressed as: Perform iterative calculations: The iteration continues until convergence, ultimately yielding distortion-free normalized coordinates. ,in, Indicates the distortion term, These represent the principal point coordinates (the projection of the optical center onto the pixel plane). These represent focal lengths (in pixels).

[0052] Combining monocular vision with prior shape constraints, 3D reconstruction of the surface image is performed, as follows: For a surface camera, given a pixel... And the corresponding undistorted ray direction, but depth If the unknown, then the prior knowledge that bearing rollers are cylindrical can be used for 3D reconstruction, assuming pixel points... In the camera coordinate system The corresponding ray direction is: ,in, This represents the inverse matrix of the intrinsic parameters of the surface camera. Represents distortion-free normalized coordinates. Let represent the ray direction vector in the camera 1 coordinate system. The parametric equation of this ray is: , Indicates the depth along the ray. In the surface camera coordinate system, the depth is... Points on the ray.

[0053] In the world coordinate system In the middle, it is assumed that the axis is parallel to The axis passes through the point. The equation for the cylindrical surface of the bearing roller is then expressed as: Here, Indicates the radius of the bearing rollers. Indicates the axis of the cylinder at A point on a plane.

[0054] Transform the ray equation to the world coordinate system Then we have: , In world coordinates, the depth is... The point on the ray will of Substituting the components into the equation of the cylindrical surface, we get: , This represents the first two rows of the rotation matrix. This represents the first two components of the translation vector, which is related to depth. quadratic equation , Let represent the coefficients of the quadratic equation with respect to depth d. Since a ray and a cylinder generally have two intersection points (entry and exit), let the depth be d. The smaller solution corresponds to the point on the cylindrical surface closer to the surface camera. The depth is then calculated. Afterwards, return That is, to obtain the points on the surface of the cylinder in the world coordinate system. Precise three-dimensional coordinates To form surface point clouds.

[0055] For contour cameras, due to their unique backlighting imaging method, specifically the position of the backlight plane in the world coordinate system... This indicates that the depth of the object point is known. For contour edge points extracted from the contour image, i.e., sub-pixel points... First, perform the same distortion correction to obtain the distortion-free coordinates of the contour camera. In the camera coordinate system In the middle, distortion-free coordinates The corresponding point is located on the ray. Up, transform to world coordinate system In the middle, there are: The third line of this equation gives information about depth. The equation can be directly solved to determine the depth. Then increase the depth Substituting back the first two lines, we can directly calculate the answer. This gives us the three-dimensional coordinates of the contour edge points in the world coordinate system. This leads to the formation of point clouds along the contour edges. The inverse matrix must be used to represent the transition from pixels to normalized coordinates.

[0056] Through the above calculations, two independent transformations to the world coordinate system were obtained. Point clouds, i.e., surface point clouds and outline edge point cloud Theoretically, they are already in the same coordinate system, allowing for coarse alignment verification based on the contour edge point cloud. The fitted cylinder radius and axis should be consistent with those based on surface point clouds. The fitted results are consistent within the tolerance range.

[0057] Furthermore, slight deviations may exist between the two due to calibration errors. Fine registration can also be achieved using the high-precision geometric features of the bearing rollers (such as end faces or specific reference surfaces). Therefore, the fine registration process can be modeled as a least-squares optimization problem to find an optimal rigid body transformation. Apply this rigid body transformation to one of the point clouds, for example This minimizes the sum of distances between the two point clouds in a common region (such as an end face or a specific reference plane). The rigid body transformation can then be expressed as follows: The optimal rigid body transformation can be found by efficiently matching the feature surfaces using the iterative nearest point algorithm or principal component analysis.

[0058] The final result is: the corrected surface point cloud. Each point contains its three-dimensional coordinates. And the associated original texture color (from ); Corrected contour 3D point cloud A high-precision set of three-dimensional edge points defines the actual geometric profile of the bearing rollers.

[0059] All data point coordinates in this step are in physical units (millimeters), allowing for direct measurement of distance, angle, area, etc.

[0060] Surface defects (in the corrected surface point cloud) (in the middle) and contour deviation (through analysis of the corrected 3D point cloud of the contour) The differences between the theoretical model and the actual model can be correlated and analyzed within the same three-dimensional space. For example, it is possible to accurately report "at an axial Z=15.2mm, the circumferential angle..." There is an oil stain at that location.

[0061] As input for subsequent steps, the PatchCore algorithm can perform more accurate neighborhood analysis based on the mapped surface texture image and its known 3D coordinates; calculus analysis, on the other hand, is performed directly on the corrected contour 3D point cloud. The process is performed on the provided precise two-dimensional contour lines. This completes the transformation from imaging sensor data to a measurable three-dimensional physical entity description, laying a solid mathematical and data foundation for subsequent intelligent analysis and precision measurement.

[0062] In one embodiment, generating the unfolded map of the surface under test and the corresponding coordinate mapping relationship based on the registered surface point cloud includes the following steps: Cylindrical fitting is performed on the registered surface point cloud, and the cylinder parameters are optimized by nonlinear least squares method to minimize the sum of squares of the deviations between the distances from each 3D point on the surface to the axis of the fitted cylinder and the fitted radius, thus obtaining the optimal cylinder parameters. The cylinder parameters include the axial direction, a point on the axis, and the radius. Based on the three-dimensional points of each surface and the optimal cylinder parameters, the cylinder coordinates corresponding to the three-dimensional points of each surface are calculated. The cylinder coordinates include axial coordinates and circumferential angle coordinates. The vertical pixel position is determined according to the axial coordinates and the preset axial pixel density. The horizontal pixel position is determined according to the circumferential angle coordinates and the total pixel width of the cylinder unfolded diagram. The circumferential boundary is then processed by wrapping around it. The size of the cylindrical unfolded pattern is determined based on the axial coordinate range and axial pixel density. The cylindrical unfolded pattern is generated by combining the color information corresponding to each horizontal and vertical pixel position. The missing areas are filled by interpolation based on the color information of neighboring pixels to obtain a continuous and complete surface unfolded pattern. A one-to-one correspondence is established between the pixels of the cylindrical unfolded image, the three-dimensional coordinates, and the cylindrical coordinates to form a coordinate mapping relationship. The coordinate mapping relationship is then used to achieve a unified positional mapping between the unfolded image pixels, the surface point cloud, and the contour edge point cloud, providing a unified spatial positioning benchmark for determining whether the surface defect area and the contour abnormal area correspond to the same physical area.

[0063] This embodiment can be understood as a mapping from a three-dimensional surface to a surface unfolded map, that is, by acquiring the corrected surface point cloud. ,in It also includes the RGB color value of each point. The cylindrical surface is parameterized and unfolded into a two-dimensional planar image, which facilitates subsequent image patch-based algorithm processing.

[0064] The specific process is as follows: Although the cylindrical assumption has been used for 3D reconstruction, to obtain accurate unfolding relationships, the surface point cloud needs to be analyzed. Perform a refit to obtain the optimal cylindrical surface parameters.

[0065] A cylindrical surface is defined by a unit vector along its axis. A point on the axis, i.e., a reference point on the axis of the cylinder. and bearing roller radius It is uniquely determined. Therefore, the parameters of the cylindrical surface are solved. This makes the surface point cloud If the sum of the squares of the distances to the cylindrical surface is minimized, then: This is a nonlinear least squares problem, which can be solved by using the Levenberg-Marquardt algorithm or by using principal component analysis (PCA) to initially estimate the axis direction and then iteratively optimizing to obtain the optimal parameters. , The coordinates of the i-th 3D point in the world coordinate system W.

[0066] For each 3D point Calculate each three-dimensional point Cylindrical coordinates on the fitted cylindrical surface axial coordinates Represents each three-dimensional point To the axis reference point The projection length is as follows: .

[0067] Angular coordinates Representing a three-dimensional point The circumferential angular position on the cross-section of the bearing roller. Specifically: calculate the three-dimensional point. Projection vector on the plane perpendicular to the axis , represented as In this plane, a reference direction vector is selected. (For example, the projection of the X-axis of the world coordinate system onto a plane), then the angle for: ,Adjustment , The projection vector of point i onto the plane perpendicular to the axis. This represents the reference direction vector within the cross-section.

[0068] Set the resolution of the unfolded image: axial pixel density Total number of pixels in the circumference (right (Radius); Axial pixel coordinates of the unfolded image: ,in, Circumferential pixel coordinates: , Width of the unfolded image (in pixels).

[0069] Create a size of blank image unfolded diagram height Each 3D point cloud color Assign to image position Since point clouds are discrete, there may be multiple points mapping to the same pixel or some pixels having no corresponding points. Therefore, bilinear interpolation or inverse distance weighting methods are used to fill the entire image with the colors of neighboring points, generating a continuous and complete surface unfolded texture map. .

[0070] Ultimately, a surface unfolded texture map will be obtained. and offline construction of pixel coordinates To three-dimensional world coordinates and cylindrical coordinates The coordinate mapping relationship LUT is represented as follows: ,in, For circumferential angle, This is the axial coordinate; during detection, the pixel coordinates on the two-dimensional image are mapped using this coordinate relationship. It directly maps to its corresponding three-dimensional spatial position and cylindrical coordinate parameters, that is, it maps each pixel on the surface unfolded texture map back to its three-dimensional coordinates and cylindrical coordinates through coordinate mapping relationship.

[0071] In one embodiment, constructing the spatial association between the surface defect region and the contour anomaly region based on the coordinate mapping relationship includes the following steps: After completing surface defect detection, the set of pixel coordinates corresponding to the surface defect region is obtained; after completing contour analysis, the set of world coordinates corresponding to the contour anomaly region is obtained. Using coordinate mapping relationships, the pixel coordinates corresponding to the surface defect area are converted into world coordinates and cylindrical coordinates, where the cylindrical coordinates include axial coordinates and circumferential angular coordinates; the world coordinates corresponding to the contour anomaly area are obtained based on the contour analysis results, and the corresponding axial coordinates and radial positions are calculated. For any surface defect region, calculate its axial distance, circumferential distance, and three-dimensional spatial distance from each contour anomaly region. The axial distance is the absolute value of the difference between the axial coordinates of the centers of the two regions; the circumferential distance is calculated based on the difference between the circumferential angular coordinates of the two regions, taking the minimum angular distance by considering the circumferential wrapping characteristics of the cylindrical surface; the three-dimensional spatial distance is calculated based on the Euclidean distance between the center points of the two regions in the world coordinate system. Based on the preset spatial tolerance, determine whether the axial distance, circumferential distance, and three-dimensional spatial distance meet the association conditions. When the axial distance, circumferential distance, and three-dimensional spatial distance are all less than the corresponding preset spatial tolerance, determine that the surface defect area and the contour abnormal area belong to the same physical area, and establish a corresponding associated defect pair; otherwise, do not establish an associated defect pair. For each established pair of associated defects, spatial consistency, positional continuity, and association reliability are calculated. Spatial consistency is calculated based on the axial distance, circumferential distance, and three-dimensional spatial distance. Positional continuity is calculated based on the continuous distribution of surface defect areas and contour anomaly areas in the axial and circumferential directions. Association reliability is determined by combining spatial consistency, positional continuity, and defect type. Spatial relationships are constructed based on the spatial consistency, locational continuity, and association credibility of each associated defect.

[0072] In one specific embodiment, the construction of the enhanced memory library based on the defect-free surface unfolding atlas includes the following steps: Feature extraction is performed on the unfolded atlas of defect-free surfaces to obtain at least two feature maps of different scales, and multidimensional feature vectors corresponding to the spatial locations of each feature map are extracted respectively; Local feature aggregation is performed on each feature vector, the statistical features of the feature vectors in the preset neighborhood are calculated, and normalization is performed to obtain normalized feature vectors. The normalized feature vectors are combined to form the feature vector set of the corresponding surface unfolded map. The feature vector sets corresponding to each surface unfolding map are collected to form a global feature pool. Cosine distance is used as the distance metric. A preset number of representative features are selected from the global feature pool according to a preset selection criterion to build an initial memory bank. The features of normal samples collected under different working conditions are added to the initial memory, and the initial memory is enhanced to obtain the enhanced memory.

[0073] In one embodiment, the defect region is identified and detected using the PatchCore algorithm. The PatchCore algorithm is divided into a training phase and an inference phase. The training phase involves building a normal feature memory offline, while the inference phase involves detecting anomalies online.

[0074] This embodiment describes the training phase, specifically the construction of a robust normal feature memory base: The training set, i.e., the surface development atlas of defect-free bearing rollers, consists of... The surface development diagrams of a defect-free, normal bearing roller constitute the surface development diagram set, which is represented as follows: , , represents the i-th surface unfolded image.

[0075] A ResNet34 dataset, jointly pre-trained on ImageNet and industrial surface datasets (such as MVTecAD), is used. The classification head is removed, and the output up to the second and third residual blocks (denoted as layer2 and layer3) is extracted as feature maps. These mid-layer features are sensitive to texture and local structural changes, making them suitable for oil stain detection. For each surface unfolded image of a defect-free bearing roller, forward propagation yields two feature maps. and Coordinates of each spatial location in each feature map Take its corresponding Ch-dimensional feature vector Extracted, This represents the Ch-dimensional original feature vector extracted at spatial location (h, w).

[0076] To enhance robustness to affine transformations, for each eigenvector... Perform local feature aggregation and take its The mean of the eigenvectors within the neighborhood is then subjected to L2 normalization to obtain the normalized eigenvectors, as shown below: , ,in, Indicates position The feature vector after aggregating the mean of the 3×3 neighborhood; All normalized feature vectors from layer 2 and layer 3 are merged to form the feature vector set for each surface unfolded map. ,in, , This represents the set of all normalized eigenvectors of a single unfolded graph I. This represents the total number of eigenvectors in a single unfolded graph.

[0077] All The feature vector set of the surface unfolded graph is collected to obtain the global feature pool. The total number of feature vectors in the global feature pool is [size missing]. ; To improve efficiency, an iterative greedy KCenter algorithm is used from... Select a subset of size M from the data. As a memory bank, its goal is to minimize the maximum coverage error: , This represents the optimal memory that minimizes the maximum coverage error. Represents the cosine distance.

[0078] Distance metric Using cosine distance: , Let represent the inner product of feature vectors f and l. Here, cosine distance is chosen because it is sensitive to the directional differences of feature vectors and can better capture texture changes caused by oil stains. To address the possibility that oil stains may be mistaken for highlights, additional sample features with normal highlights collected under specific lighting angles are added to the core set to increase the coverage of normal gloss changes in the memory bank, forming a memory bank enhancement and reducing false alarms.

[0079] In one embodiment, the steps of extracting the characterization features of the unfolded surface of the test surface, determining the surface defect region, recovering the corresponding spatial position through coordinate mapping, calculating the defect features, and obtaining the surface defect result include the following steps: The characterization features of the surface unfolded map to be tested are extracted and similarity calculation is performed with the augmented memory. The nearest neighbor distance between each feature to be tested and the corresponding feature in the augmented memory is used as the anomaly score of the corresponding region. Each anomaly score is mapped to the spatial location of the corresponding feature map to obtain anomaly score maps at different scales. Each anomaly score map is upsampled to the same size as the original unfolded map, and then fused by taking the maximum value at the pixel level to generate an anomaly heatmap; the segmentation threshold is adaptively determined based on the anomaly heatmap to generate a binary defect mask. Connected components are extracted based on a binary defect mask, and regions that do not meet preset conditions are filtered out to obtain the target defect region; By using coordinate mapping relationships, the pixel coordinates corresponding to the target defect area are restored to the world coordinate system and cylindrical coordinate system. The spatial location, geometric features and texture features of the target defect area are calculated, the defect type is identified, and the surface defect result is formed.

[0080] This embodiment is the reasoning stage, in which anomaly detection and localization are performed on each defect area. The specific process is as follows.

[0081] First: The unfolded diagram of the test surface of the bearing roller under test. Perform the following steps: Feature extraction and similarity calculation: Extract a set of multi-scale local feature vectors in the same manner as the training phase. For each test feature Calculate the nearest neighbor cosine distance from it to memory C, and use it as the anomaly score for that local region. : , This represents the set of multi-scale locally normalized feature vectors of the test image. This represents the total number of test feature vectors. Let t represent the normalized feature vector of the t-th test, where t=1,…, , This represents the anomaly score of the t-th local region.

[0082] Second: Based on the abnormal scores, perform the following steps: Each feature vector Abnormal scores Mapping back to spatial coordinates on its source feature map Obtain the anomaly score maps for layer 2 and layer 3. ; The abnormal score map Upsampled to the original unfolded image respectively Same size The upsampling score map is obtained. The two score maps at different scales are fused by taking the maximum value at the pixel level to generate the final anomaly heatmap A. The maximum value operation helps to preserve significant anomalies detected at any scale, generating the final anomaly heatmap.

[0083] Third: Post-processing and defect segmentation of abnormal heat maps, including the following steps: Calculate the Otsu threshold of the anomalous heatmap A And set the final threshold. , For example, 1.2, to improve the recall rate for oil stains). Generate a binary defect mask B: ; A closing operation (dilation followed by erosion, using a circular structuring element) is performed on the binarized defect mask B to connect adjacent outliers. Then, an opening operation (erosion followed by dilation) is performed to remove isolated noise points. The processed binary image is denoted as [image 1]. ; In binary image The search is performed on the connected components, and each connected component corresponds to a defect region. Filter out areas smaller than the threshold The defect area is obtained from the tiny area (noise). This represents the i-th defect region (connected region). This represents the area filtering threshold; connected components with pixel areas smaller than this value are considered noise and are removed.

[0084] Fourth: Perform three-dimensional localization and quantification of the defect area, including the following steps: For each defect area Using coordinate mapping relationships (LUTs), its pixel coordinate set is obtained. Mapping back to a set of three-dimensional world coordinates and cylindrical coordinate set , This represents the cylindrical coordinates (angle and axial height) corresponding to the j-th pixel. Represents the unfolded pixel coordinates of the j-th pixel within the defect region; The geometric and textural features of the computational defect include: 3D center location: ; Axial range is ; circumferential range is (Note the handling of angle periods); Approximate area: On a 3D surface, the actual surface area is estimated by triangulation of the projected region or by using the Jacobian determinant (conversion from pixel area to surface area); Average anomaly score: , Let A represent the i-th defect region (connected region), and let A represent the abnormal heatmap.

[0085] In a specific embodiment, it is necessary to describe in detail: geometric features include a three-dimensional center, distribution range, and approximate area; texture features include average anomaly score, contrast, and correlation, then: The coordinates of the three-dimensional center of the defect region can be obtained by using the three-dimensional world coordinates of each pixel within the defect region. The distribution range of the defect region along the axial direction is obtained by using the extreme values ​​of the axial coordinates in the cylindrical coordinate set; By using the extreme values ​​of the angular coordinates in the cylindrical coordinate set and performing continuity processing on the angular period when the angular coordinates cross the period boundary, the distribution range of the defect region in the circumferential direction can be obtained. From the geometry of the defect region, eccentricity and elongation are extracted as shape features; By using the pixel coordinates of the projection region of the defect region onto the three-dimensional surface, the pixel area is converted into the surface area, and the approximate area of ​​the defect region is obtained. The average anomaly score of the defective region is obtained by using the anomaly scores of each pixel within the defective region. Extract the average RGB value as a color feature from the RGB values ​​of the defect area in the test image; Saturation S and brightness V are extracted as color features from the pixel values ​​of the defective region in HSV space. A gray-level co-occurrence matrix is ​​constructed based on the pixel gray-level values ​​and spatial location relationships within the defect area, and contrast and correlation are extracted based on the gray-level co-occurrence matrix. Contextual features are extracted based on the average anomaly score of the defective region and its difference from the score of the surrounding normal region.

[0086] In step S500, the defect area is identified based on a classifier to obtain surface defect results, including at least the defect category and its corresponding confidence level. This classifier is trained using geometric features, texture features, color features, shape features, and contextual features. Specifically, a classifier (such as a Support Vector Machine, SVM) is trained based on the feature vectors of historically labeled samples to distinguish "oil stains" from other defects (such as scratches and dents). The feature vectors include geometric features, texture features, color features, shape features, and contextual features; color features: the defect area is within... The average RGB values, saturation (S) and luminance (V) in HSV space; texture features: contrast and correlation of the gray-level co-occurrence matrix (GLCM) within the defect region; shape features: eccentricity and elongation of the region; contextual features: average anomaly score of the region. The score of each defective region is compared with the score of the surrounding normal region; based on this classifier, each defective region is analyzed. Predict its defect category and confidence level .

[0087] Ultimately, a surface defect result can be generated for each bearing roller under test. The surface defect result includes at least the defect category and the corresponding confidence level, and may also include some information generated beforehand, such as visualization results and summary indicators. The defect category, its corresponding confidence level, and its location information are used to form a list of each detected defect area. A list of defects, such as: defect categories (e.g., "oil stain") and confidence level ; Three-dimensional center position and cylindrical coordinates Axial and circumferential extension range; estimated surface area and average anomaly score. Let represent the average circumferential angle of the i-th defect region.

[0088] The visualization results include unfolded diagrams with labeled defect areas. and its corresponding anomalous heatmap A; The summary indicators include the total number of defects, the number of oil stain defects, and the largest defect area.

[0089] This report, along with the contour analysis results, will be used to comprehensively assess the workpiece's quality. By combining 3D mapping, the location of surface defects can be spatially correlated with the location of contour deviations, providing richer information for process diagnostics.

[0090] In a specific embodiment, the actual profile curve of the bearing roller is reconstructed based on the registered profile edge point cloud, the profile deviation and variation characteristics between the actual profile curve and the theoretical profile curve are calculated, the profile anomaly region and its corresponding world coordinates are determined, and the profile analysis result is formed, including the following steps: Based on the registered contour edge point cloud, the axis and axis direction are determined, the radial distance corresponding to each contour point is calculated, and the points are sorted according to the axial coordinates to form an ordered contour point set; outlier points are removed by the median absolute deviation filtering method to obtain the contour data after gross error removal. Based on the contour data, a discrete contour curve is constructed, and a smooth spline is used for fitting. An objective function including a fitting error term and a smoothing constraint term is constructed, and the actual contour curve is obtained by optimization. The scaling factor, axial translation, and radial translation between the actual contour curve and the theoretical contour curve are optimized using the least squares method to eliminate the rigid displacement error between the two and obtain the registered actual contour curve. The contour deviation, derivative characteristics, and curvature characteristics between the registered actual contour curve and the theoretical contour curve are calculated, and multi-level contour analysis is performed based on the contour deviation, derivative characteristics, and curvature characteristics. Based on the results of multi-level contour analysis, abnormal contour regions are identified, and the world coordinate positions corresponding to the abnormal contour regions are determined. Contour defect diagnostic information is formed by combining the continuity, change trend and distribution characteristics of the abnormal contour regions. By combining contour accuracy indicators and contour defect diagnosis information, contour analysis results are generated.

[0091] This step involves a detailed analysis of the correction contour edge point cloud, converting it into a parametric contour curve, thus obtaining the corrected 3D contour point cloud. ,in, These points are theoretically located near the intersection of the cylindrical surface of the bearing roller and the end face (or the measurement reference surface), representing the outer edge contour of the bearing roller. Finally, an ordered, parameterized two-dimensional contour curve representing the generatrix of the bearing roller is extracted from the three-dimensional point cloud.

[0092] In a specific embodiment, this can be achieved through the following steps: First, the point cloud of the corrected contour edge is acquired and preprocessed to obtain the parameterized actual contour curve, including the following steps: Step 1: Construct a coordinate system and project the point cloud of the corrected contour edge. Although a cylindrical model has been used, to obtain the highest accuracy axes for contour analysis, the 3D point cloud of the corrected contour is then... We fit a straight line in space using total least squares (TLS). Let the unit direction vector of the axis be... A point on the axis, i.e., the reference point of the axis, is The optimization problem is: Its analytical solution is: the centroid of the point cloud As covariance matrix The unit eigenvector corresponding to the smallest eigenvalue is Define the axial coordinates of the i-th contour point as follows: , This represents the unit direction vector of the fitted axis.

[0093] To obtain a single-sided generatrix, the point cloud needs to be divided into "upper" and "lower" sides (corresponding to the two generatrixes of a cylinder). The radial vector of the i-th contour point relative to the axis is calculated, denoted as... Choose a reference direction (such as the reference direction vector of the world coordinate system's X-axis on a plane perpendicular to a). , ) Calculate the dot product symbol . with dot product symbol Points that are consistent (e.g., positive) are grouped into a set of points on one side of the generatrix. According to Arrange in ascending order to obtain an ordered list of points. ,in, This represents the axial coordinate of the k-th point in a single-sided busbar point sequence. This represents the corresponding radial vector. Indicates the total number of points on the busbar. This represents the world coordinates of the i-th contour point.

[0094] radial distance Due to measurement noise and potential outliers, a filter based on median absolute deviation (MAD) is employed: the sequence is calculated. the median of and absolute deviation sequence The median MAD, excluding those that satisfy The outliers were identified, and the radial distance sequence after gross error removal was obtained.

[0095] The second step is to parameterize and smooth the contour curve. Since discrete points contain noise, and subsequent calculus analysis requires a differentiable function, a smoothing function is needed to fit the data. Specifically, the axial coordinate z is selected as the parameter, and the fitting function is... Represents radial distance. Domain is... The data was fitted using penalized least squares splines (smooth splines). The goal is to find the function. Minimize the objective function: ,in, The fitting weight for the k-th data point (can be set according to measurement uncertainty). It is a smoothing parameter that controls the balance between goodness of fit and smoothness. Indicates the total number of points on the busbar. This represents the covariance matrix, and the curvature of the penalty function ensures a smooth curve. Indicates the smoothing parameter. The solution can be automatically selected using generalized cross-validation (GCV), and its solution is a natural cubic spline function with continuous second derivatives. Solving for the smooth function yields the smooth function. and its first derivative and second derivative The spline expression is the mathematical representation of the actual measured profile, and the actual profile curve is obtained at this time.

[0096] Secondly, after obtaining the actual contour curve, it needs to be registered with the theoretical contour curve to obtain the registered actual contour curve. The theoretical design model of the logarithmic generatrix is ​​known, and the Lundberg logarithmic curve can be used, its general form being: ,in, Indicates the theoretical datum radius of the cylinder. Represents the logarithmic correction term. This represents a vector of design parameters; for example, a common form is: ,in, Represents the convexity coefficient. It is the effective length. It is the central location. This represents the axial coordinate of the k-th contour point. Represents axial coordinates. First and second derivatives of the theoretical model. It can be solved analytically.

[0097] Before registration, it is necessary to eliminate rigid displacement errors (alignment errors and radius deviations) between the actual and theoretical profiles. Let the registration transformation be: ,in, It is the scaling factor (usually close to 1). It is the axial translation amount. It is the radial translation amount. This represents the actual profile curve after registration.

[0098] The optimal parameters are obtained by least squares fitting. , This is a nonlinear optimization problem, solvable using Gauss-Newton's method or the Levenberg-Marquardt algorithm. The actual contour function is updated using the registration parameters: Its derivative needs to be adjusted accordingly: The scaling factor, axial translation, and radial translation are optimized by the least squares method to eliminate the rigid displacement error between the actual contour and the theoretical contour, thus obtaining the actual contour curve after registration.

[0099] Based on the actual profile curve after registration, multi-level comparison analysis is performed with the theoretical profile curve of the bearing roller to obtain profile accuracy index and defect diagnosis information. In order to obtain more accurate analysis results, this embodiment adopts multi-level comparison analysis, which includes at least zero-order deviation, first derivative and slope analysis, second derivative and slope analysis, convexity analysis and local defect analysis.

[0100] Zero-order analysis includes comparing profile shape, profile deviation function, key indicators, maximum positive deviation, maximum negative deviation, peak-to-peak deviation, mean absolute deviation, and root mean square deviation. First derivative analysis: Evaluate the trend change of the contour: Obtain the aligned actual contour curve and theoretical contour curve; Calculate the first derivative of the actual contour curve and theoretical contour curve along the axial position to obtain the actual slope and theoretical slope. At each axial position, the difference between the actual slope and the theoretical slope is calculated to obtain the slope deviation; The actual slope and the theoretical slope are respectively subjected to arctangent calculation to obtain the actual tangent inclination angle and the theoretical tangent inclination angle; the tangent inclination angle is the angle between the tangent of the contour curve and the reference direction, and the angle unit is kept uniform; At each axial position, the difference between the actual tangent inclination angle and the theoretical tangent inclination angle is calculated to obtain the tangent inclination angle deviation; For the slope deviation sequence at all axial positions, the maximum, minimum, and root mean square values ​​are calculated; for the tangent inclination angle deviation sequence at all axial positions, the maximum absolute value is calculated, which serves as a quantitative evaluation index for profile consistency.

[0101] It also includes second-derivative analysis: calculating the actual curvature and theoretical curvature and their deviation, and diagnosing local machining defects based on the maximum absolute value of the rate of change of curvature; Curvature is the most sensitive indicator for evaluating profile accuracy and is directly related to contact stress; Obtain the actual contour curve and the theoretical contour curve; calculate the first and second derivatives of the actual contour curve and the theoretical contour curve respectively, and substitute them into the curvature formula to calculate the actual curvature and the theoretical curvature. At each axial position, the difference between the actual curvature and the theoretical curvature is calculated to obtain the curvature deviation; The first derivatives of the actual curvature and the theoretical curvature along the axial position are obtained to obtain the rate of change of the actual curvature and the rate of change of the theoretical curvature. Statistical analysis is performed on the actual curvature at all axial positions to determine its maximum and minimum values ​​and corresponding axial positions; statistical analysis is performed on the curvature deviation at all axial positions to calculate its peak-to-peak value and root mean square value; and the maximum absolute value of the rate of change of curvature at all axial positions is calculated. It also includes convexity and support length analysis: calculating the radial convexity of the line connecting the center point of the profile to the two endpoints and the position of the maximum radial offset, and statistically analyzing the ratio of the axial support length of the actual profile to the theoretical profile within the specified deviation zone; Obtain the effective range of the contour curve, and determine the axial position of the start point, the axial position of the end point, and the axial position of the center point of the effective range; Obtain the radial coordinates of the starting point and the ending point, and construct the reference line equation between the two endpoints based on the radial coordinates of the starting point and the ending point; At the axial position of the center point, calculate the difference between the radial coordinates of the profile and the radial coordinates of the reference line to obtain the convexity of the center point; or within the effective range, calculate the difference sequence between the radial coordinates of the profile and the radial coordinates of the reference line point by point, extract the maximum value in the difference sequence and its corresponding axial position to obtain the maximum radial offset and the offset position. Set a deviation zone threshold, calculate the ratio of the axial length of the actual contour and the theoretical contour within the deviation zone threshold range to the total length of the effective interval, and obtain the contour support length ratio. It also includes local defect detection: marking local abnormal regions based on the rate of curvature change threshold, and calculating the second derivative of the deviation function to identify abrupt changes in the contour shape; Detecting small local anomalies using higher-order derivatives: Set a curvature change rate threshold; compare the curvature change rate with the threshold, and mark the axial position region where the absolute value of the curvature change rate exceeds the threshold; Obtain the contour deviation function obtained in the aforementioned steps; calculate the second derivative of the deviation function along the axial position to obtain the deviation curvature; extract the local extreme points of the deviation curvature and their corresponding axial positions as indicators of abrupt changes in the contour shape.

[0102] Zero-order deviation includes at least a direct comparison of profile shape and key indicators; the direct comparison of profile shape is calculated using the profile deviation function, expressed as: ,in, Represents the profile deviation function. This represents the actual profile curve after registration. This represents the theoretical profile curve; key indicators include maximum positive deviation, maximum negative deviation, peak-to-peak deviation, mean absolute deviation, and root mean square deviation. The maximum positive deviation is expressed as: The maximum negative deviation is expressed as: Peak-to-peak deviation, expressed as: The mean absolute deviation is expressed as: Root mean square deviation, expressed as: ,in, , This represents the axial span of the profile, i.e., the integral normalized length.

[0103] First-order derivative analysis, also known as slope and tangent angle analysis, is used to assess the trend changes of a profile. The actual slope function is expressed as: The theoretical slope function is expressed as: The slope deviation is expressed as: The actual tangent angle is expressed as: ; Tangent angle deviation is expressed as: (Unit: radians or degrees); Key indicators are: Maximum value, minimum value ; The maximum absolute value.

[0104] Second derivative and curvature analysis are the core of comparative analysis. Curvature is the most sensitive indicator for evaluating profile accuracy and is directly related to contact stress. Planar curves The curvature formula is expressed as: ,in, The ordinate of the plane curve in the derivation of the curvature formula is represented. These represent the first and second derivatives of the registered contour, respectively.

[0105] Regarding the bearing roller profile (The slope is very small), therefore the curvature formula can usually be simplified to However, in this embodiment, for more accurate calculations, the complete formula is used to calculate the actual curvature calculated by the curvature function. and theoretical curvature Based on actual curvature and theoretical curvature The curvature deviation is obtained: ; The rate of change of curvature is obtained by using the derivative of curvature, and the smoothness of curvature is evaluated by the following formula: Then the actual curvature derivative can be calculated. and theoretical curvature derivative .

[0106] This allows us to determine key curvature parameters: maximum and minimum curvature values ​​and their locations; and curvature deviation. Peak-to-peak value, RMS; maximum absolute value of the rate of change of curvature. Local mutations (which may correspond to machining vibration, grinding wheel wear, or local defects) can be diagnosed by the maximum absolute value of the rate of change of curvature.

[0107] Convexity is defined as the radial distance between the lines connecting the center point and the two endpoints of the profile. Let the effective interval be... The center point is The equation of the line connecting the endpoints is: Then the convexity is expressed as: To make the calculation results more accurate, we can find the maximum radial offset, expressed as: And record the axial position corresponding to the maximum convexity. , This indicates the maximum convexity, which is the maximum radial offset within the effective range. Indicates the endpoints of the effective interval for convexity calculation. Indicates the center point of the valid interval. This represents the equation of the line connecting the endpoints.

[0108] In bearing design, the "support length" or "camber extension length" of the profile is also crucial. It's also possible to calculate the deviation zone between the actual and theoretical profiles (e.g., ...). The percentage of axial length within ).

[0109] Finally, higher-order derivatives are used to detect anomalies in small local defects, including setting a threshold for the rate of curvature change. The marker satisfies These areas may correspond to scratches, dent edges, and other areas; for the deviation function Calculate its second derivative , The local extreme points indicate the inflection points of the deviation curve, which may correspond to abrupt changes in the contour shape.

[0110] Additionally, frequency domain analysis can be performed. Frequency domain analysis is used for process diagnostics and is not a mandatory step. (Regarding the deviation function...) Perform a Fourier transform and analyze its spectrum: The peak values ​​in the spectrum may correspond to specific periodic errors (such as machine tool spindle rotation error or feed screw error), providing diagnostic information for process debugging.

[0111] Following the comparative analysis, the uncertainty arising from error propagation is calculated, and the profile analysis results are then derived from the uncertainty assessment. Due to measurement noise and fitting errors, all quantitative indicators require uncertainty assessment. This is achieved using the Monte Carlo method or the error propagation law.

[0112] Based on the measurement uncertainty of point cloud coordinates and the covariance matrix of smooth spline fitting, the standard uncertainty of function values, derivatives, and profile accuracy indices is calculated using the Monte Carlo method or error propagation law to obtain the uncertainty assessment results and form the profile analysis results. This includes the following steps: Based on the Monte Carlo method or error propagation law, the standard uncertainty of each evaluation index is calculated according to the measurement uncertainty of point cloud coordinates and the fitted covariance matrix. Due to measurement noise and fitting error, all quantification metrics must be accompanied by uncertainty assessments, using the Monte Carlo method or error propagation law. Input uncertainty: The measurement uncertainty of the point cloud coordinates is obtained as the input uncertainty. The measurement uncertainty of the point cloud coordinates is determined by the camera calibration and sub-pixel extraction accuracy.

[0113] Parameter fitting uncertainty: During the smooth spline fitting process, the covariance matrix of the fitting parameters is calculated; based on the covariance matrix, the uncertainty band of the fitted function value and its derivative is calculated. In the smooth spline fitting, the uncertainty band of the function value and its derivative is calculated using the covariance matrix.

[0114] Indicator uncertainty: Calculate the standard uncertainty of each quantitative evaluation indicator obtained in the previous steps using the error propagation law or the Monte Carlo method, and calculate the standard uncertainty of the indicator through error propagation.

[0115] Specifically, this involves: obtaining the measurement uncertainty of point cloud coordinates. The measurement uncertainty here is determined by camera calibration and sub-pixel extraction accuracy; in smooth spline fitting, the uncertainty band of the function value and derivative is calculated using the covariance matrix; through error propagation, the uncertainty is calculated... Standard uncertainty of indicators such as [missing information].

[0116] After the above process is completed, a structured contour analysis result will be obtained, including: Registration results: offset parameters and its uncertainty; Overall deviation index: and uncertainty; First derivative index: ; Curvature index: ; convexity parameter or and its location Uncertainty; Local anomaly list: Location, amplitude, and possible type of detected curvature abrupt changes; Goodness of fit: coefficient of determination ; Spectral characteristics (e.g., analysis): main frequency components and amplitudes.

[0117] And visualized data curves (plottable), such as: actual contour and theoretical contour; deviation curve, actual contour and theoretical contour; deviation curve, curvature curve and theoretical curvature; curvature deviation curve; key indicator plots with confidence intervals.

[0118] The contour analysis results, along with the surface defect results, are input into a multi-level decision model to complete a comprehensive judgment of the workpiece. Through the calculus analysis in this step, the system not only provides a "qualified / unqualified" judgment but also offers a detailed "diagnostic report," indicating the specific pattern, location, and magnitude of the contour error, thus providing accurate data input for process closed-loop control.

[0119] In one embodiment, the step of constructing a spatial correlation between surface defect regions and contour anomaly regions based on coordinate mapping, performing cross-validation and fusion analysis on surface defect results and contour analysis results based on the spatial correlation, and completing quality judgment through a multi-level decision model includes the following steps: Obtain surface defect results, spatial correlation results, and contour analysis results; A multi-level decision-making model is constructed, which includes a surface defect determination unit, a contour quality determination unit, and a fusion determination unit. The surface quality evaluation result is calculated by the surface defect judgment unit based on the defect type, defect area range and defect confidence level; the contour quality evaluation result is calculated by the contour deviation, derivative characteristics, curvature characteristics and contour abnormal area. Based on the spatial correlation, the fusion judgment unit performs cross-validation and fusion analysis on the surface quality evaluation results and the contour quality evaluation results, and makes a comprehensive judgment by combining the defect type, spatial location and severity of the associated defects to obtain the final quality judgment result of the bearing roller under test.

[0120] Obtain surface defect results, including a list of defects for each defect region. Attribute tuples: ,in, Indicates the defect category (such as "oil stain" or "scratches"). Let i be the classification confidence level for the i-th defect. To estimate the surface area, The average abnormal score, Let each represent the coordinate components of the i-th defect center in the world frame. Let these represent the cylindrical coordinates of the i-th defect center. These represent the axial ranges of the i-th defect region, respectively. Let represent the circumferential width of the i-th defect region, the estimated surface area of ​​the i-th defect region, and the average anomaly score of the i-th defect region, respectively.

[0121] Obtain the contour analysis results, including: Overall deviation index: ; Curvature index: ; Convexity parameters: (Value and uncertainty); Local anomaly list: ; Constructing a spatial correlation between surface defects and contour anomalies, for defect regions and contour outliers Calculate axial distance and angle difference , and Let represent the axial coordinates and circumferential angles of the j-th local contour anomaly point, respectively. and , (These are axial and circumferential correlation tolerances, respectively), then they are marked as correlation pairs. This may indicate two manifestations of the same physical defect.

[0122] The multi-level decision model employs a rule-based multi-level decision tree, combining deterministic thresholding and fuzzy comprehensive evaluation, and sets a tolerance threshold vector. ,in, Includes surface-related thresholds, Includes contour-related thresholds.

[0123] Surface quality is determined using a multi-level decision model, with the surface defect determination unit set as a single-point area threshold, which is related to the defect category. For each defect region... The formula for calculating the single-point exceedance indicator is as follows: ; And calculate the overall surface mass fraction , is represented as: in, The normalization constant is These are the weighting coefficients. This is the threshold for the total number of defects. This is the threshold for the total defect area. This is a set of fatal defect categories (such as "cracks"). This represents the total number of defects (the total number of defective areas). This represents the total defect area (the sum of the areas of all defective regions).

[0124] The surface determination result is expressed as: in, The surface quality pass threshold (e.g., 0.8) is defined in the above formula as: total defect quantity threshold; total defect area threshold; and set of fatal defect categories (e.g., "cracks").

[0125] The profile quality is assessed using a profile quality assessment unit, and measurement uncertainty must be considered in the assessment. A guard band principle is adopted for each profile index. (such as convexity) Its standard uncertainty is Given a inclusion factor (generally =2), then the expanded uncertainty is: Each contour indicator The tolerance zone is .

[0126] Define the criteria for exceeding the outline index as follows: ; For convexity : ; For peak-to-peak deviation The sign indicating an excess is: , , These are the minimum and maximum allowable values ​​for convexity, respectively. Indicators indicating excessive convexity The threshold value for determining peak-to-peak value; Profile quality score The calculation formula is as follows: ,in, For the target value, For tolerance half width, 1 represents the weight, and 1 represents the indicator function. This is an indication that the root mean square deviation exceeds the standard. This indicates that the rate of change of curvature exceeds the standard. This is the inclusion factor.

[0127] The contour determination result is: .

[0128] Finally, after comprehensively judging the fused surface and contour results, and considering the associated information, we have: Among them, the association flag Represented as: This means that the combination of severity levels of associated defects exceeds a threshold. At that time, a comprehensive failure was triggered. This indicates the final comprehensive judgment result. Indicates the surface determination result. This indicates the contour determination result. Let i represent the estimated surface area of ​​the i-th defect. This represents the area normalization constant. This represents the amplitude normalization constant. This indicates the threshold for determining association.

[0129] Final confidence level It is calculated using the following formula: ,in, These are the calibration parameters, This is the joint probability term where all profile indicators are within the acceptable range. Let be the classification confidence score for the i-th defect.

[0130] The following section provides detailed technical effect diagrams based on a specific dataset. Please refer to the attached diagrams for further details. Figure 2 -Appendix Figure 9 As shown, Figure 2 An overall flowchart is provided. Figures 3-8 A technical flowchart for each step is provided. Figure 3 For image acquisition and 3D reconstruction; Figure 4 Registration of point clouds; Figure 5 It involves generating the unfolded diagram and constructing the coordinate mapping relationship; Figure 6 To construct and enhance a memory bank based on defect-free samples; Figure 7 For defect detection and identification; Figure 8 For contour accuracy analysis; Figure 9 It is a quality assessment.

[0131] The surface datasets used in the experiment are shown in Table 1.

[0132] Table 1 shows the surface dataset. Table 1 shows the test object and hardware parameters. The test object is a GCr15 bearing steel roller with a geometric dimension of Ø8mm×12mm and a surface roughness Ra≤0.2μm. Image acquisition is performed using a 4096×3000 resolution industrial camera with 2× telecentric lenses. The exposure time is set to 3ms to balance high resolution and anti-shake requirements of the production line.

[0133] The dataset used to train the classifier is shown in Table 2.

[0134] Table 2 shows the dataset used to train the classifier. Table 2 shows a total of 8350 samples in the dataset. Among them, 5000 samples are defect-free as the normal baseline; the defect samples include 1200 scratches, 800 indentations, 600 pits, 450 cracks and 300 burns. The five types of defects cover the common surface failure modes of bearing rollers, and the distribution of categories takes into account the model training and generalization capabilities.

[0135] The surface point cloud dataset is shown in Table 3.

[0136] Table 3 shows the surface point cloud dataset. Table 3 presents the surface point cloud dataset, recorded in the world coordinate system, including the three-dimensional coordinates of Point_ID, Xw, Yw, and Zw, as well as the intensity value. Yw is approximately 3.99 mm, close to the roller radius of 4 mm, indicating that Yw is the radial coordinate. Xw and Zw correspond to the circumferential and axial directions, respectively. The intensity value of 127-132 represents the surface reflection intensity, which is used to assist in texture feature extraction.

[0137] The contour edge point cloud dataset is shown in Table 4.

[0138] Table 4 shows the contour edge point cloud dataset. Table 4 shows the contour edge point cloud dataset. Only the Edge_ID and Xw / Yw / Zw three-dimensional coordinates are retained for the contour edge points, without the intensity column. Yw is stable at 4.001-4.003 mm, Xw increases in increments of 0.01 mm, and Zw varies slightly to facilitate contour fitting and geometric tolerance calculation.

[0139] The final experimental results are shown in Table 5.

[0140] Table 5 shows the final experimental results. Table 5 shows the final experimental results. The system's defect detection accuracy is 98.7%, the missed detection rate is 0.8%, and the false detection rate is 1.1%. In terms of geometric measurement, the point cloud registration error is 6μm and the contour analysis error is 0.5μm, which meets the submicron level accuracy requirements. The single roller full-process inspection takes 1.8s, which has the engineering feasibility of online batch inspection.

[0141] Example 2: A vision-based intelligent inspection system for bearing roller surfaces and contours, such as... Figure 10 As shown, it includes: A generation module 100 is constructed to build a world coordinate system, simultaneously acquire surface images and two-dimensional contour images of the bearing rollers at the same location, and preprocess them respectively to obtain surface point clouds and contour edge point clouds in the world coordinate system. The alignment and registration module 200 performs fitting, alignment and registration on the surface point cloud and the contour edge point cloud respectively to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. The mapping generation module 300 generates the unfolded map of the surface to be tested and the corresponding coordinate mapping relationship based on the registered surface point cloud, and keeps the surface detection process and the contour detection process using the same spatial coordinate reference. The identification module 400 is determined to construct an enhanced memory library based on the defect-free surface unfolding atlas, extract the characterization features of the surface unfolding atlas to be tested, determine the surface defect region, recover the corresponding spatial position through coordinate mapping relationship and calculate the defect features to obtain the surface defect result; Module 500 is defined to reconstruct the actual contour curve of the bearing roller based on the registered contour edge point cloud, calculate the contour deviation between the actual contour curve and the theoretical contour curve, determine the contour anomaly area and the corresponding world coordinates, and form the contour analysis result. The fusion analysis module 600 is used to construct the spatial correlation between surface defect areas and contour abnormal areas based on coordinate mapping relationships. The spatial correlation is used to determine whether surface defect areas and contour abnormal areas correspond to the same physical area, and only the surface defect results and contour analysis results corresponding to the same physical area are cross-validated and fused. The comprehensive quality judgment is completed through a multi-level decision model.

[0142] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0143] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] It should be noted that: The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0148] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

Claims

1. A vision-based intelligent detection method for the surface and contour of bearing rollers, characterized in that, Includes the following steps: A world coordinate system is constructed, and surface images and two-dimensional contour images of the bearing rollers at the same location are acquired simultaneously. These images are then preprocessed to obtain surface point clouds and contour edge point clouds in the world coordinate system. The surface point cloud and the contour edge point cloud are fitted, aligned and registered respectively to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. The surface unfolding diagram and corresponding coordinate mapping relationship of the surface to be tested are generated based on the registered surface point cloud, and the surface detection process and the contour detection process use the same spatial coordinate reference. An enhanced memory library is constructed based on a defect-free surface unfolding atlas. The characterization features of the unfolding map of the surface to be tested are extracted, the surface defect region is determined, the corresponding spatial position is recovered through coordinate mapping relationship, and the defect features are calculated to obtain the surface defect result. Based on the registration contour edge point cloud, the actual contour curve of the bearing roller is reconstructed, the contour deviation between the actual contour curve and the theoretical contour curve is calculated, the contour abnormal area and the corresponding world coordinate are determined, and the contour analysis result is formed. Based on the coordinate mapping relationship, a spatial correlation relationship is constructed between the surface defect region and the contour abnormal region. The spatial correlation relationship is used to determine whether the surface defect region and the contour abnormal region correspond to the same physical region. Cross-validation and fusion analysis are only performed on the surface defect results and contour analysis results that correspond to the same physical region. The comprehensive quality judgment is completed through a multi-level decision model. The step of generating the unfolded map of the surface to be tested and the corresponding coordinate mapping relationship based on the registered surface point cloud includes the following steps: Cylindrical fitting is performed on the registered surface point cloud, and the cylinder parameters are optimized by nonlinear least squares method to minimize the sum of squares of the deviations between the distances from each 3D point on the surface to the axis of the fitted cylinder and the fitted radius, thus obtaining the optimal cylinder parameters. The cylinder parameters include the axial direction, a point on the axis, and the radius. Based on the three-dimensional points of each surface and the optimal cylinder parameters, the cylinder coordinates corresponding to the three-dimensional points of each surface are calculated. The cylinder coordinates include axial coordinates and circumferential angle coordinates. The vertical pixel position is determined according to the axial coordinates and the preset axial pixel density. The horizontal pixel position is determined according to the circumferential angle coordinates and the total pixel width of the cylinder unfolded diagram. The circumferential boundary is then processed by wrapping around it. The size of the cylindrical unfolded pattern is determined based on the axial coordinate range and axial pixel density. The cylindrical unfolded pattern is generated by combining the color information corresponding to each horizontal and vertical pixel position. The missing areas are filled by interpolation based on the color information of neighboring pixels to obtain a continuous and complete surface unfolded pattern. A one-to-one correspondence is established between the pixels of the cylindrical unfolded image, the three-dimensional coordinates, and the cylindrical coordinates to form a coordinate mapping relationship. The coordinate mapping relationship is then used to achieve a unified positional mapping between the unfolded image pixels, the surface point cloud, and the contour edge point cloud, providing a unified spatial positioning benchmark for determining whether the surface defect area and the contour abnormal area correspond to the same physical area. The construction of the enhanced memory library based on the defect-free surface unfolding atlas includes the following steps: Feature extraction is performed on the unfolded atlas of defect-free surfaces to obtain at least two feature maps of different scales, and multidimensional feature vectors corresponding to the spatial locations of each feature map are extracted respectively; Local feature aggregation is performed on each feature vector, the statistical features of the feature vectors in the preset neighborhood are calculated, and normalization is performed to obtain normalized feature vectors. The normalized feature vectors are combined to form the feature vector set of the corresponding surface unfolded map. The feature vector sets corresponding to each surface unfolding map are collected to form a global feature pool. Cosine distance is used as the distance metric. A preset number of representative features are selected from the global feature pool according to a preset selection criterion to build an initial memory bank. The features of normal samples collected under different working conditions are added to the initial memory, and the initial memory is enhanced to obtain the enhanced memory.

2. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The surface point cloud and contour edge point cloud are obtained through the following steps: A world coordinate system is constructed by taking the ideal axial direction of the bearing rollers as the Z-axis and the radial plane as the XY plane. In response to the triggering of the bearing roller reaching the preset detection position, the surface imaging station and the contour imaging station are synchronously controlled to acquire surface images and two-dimensional contour images respectively. Select the optimal surface image, convert the coordinates of each pixel in the optimal surface image into the initial distortion-free normalized coordinates in the world coordinate system, and perform correction processing to obtain the corrected image coordinates; Based on the corrected image coordinates, the corresponding ray direction, and the prior model of the cylinder, the intersection point of the ray and the cylinder surface is calculated, and the intersection point closest to the camera is selected as the corresponding surface point. The three-dimensional coordinates of the surface point in the world coordinate system are obtained to form a surface point cloud, where the bearing roller is a cylinder. A two-dimensional contour image is acquired, sub-pixel edges are extracted and distortion correction is performed to obtain distortion-free coordinates. The distortion-free coordinates are then transformed to the world coordinate system using camera intrinsics to obtain the three-dimensional coordinates of the contour edge points, thus forming a contour edge point cloud.

3. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The process of fitting, aligning, and registering the surface point cloud and the contour edge point cloud to obtain the registered surface point cloud and the registered contour edge point cloud includes the following steps: Cylindrical fitting is performed on the surface point cloud and the contour edge point cloud respectively to obtain the radius and axis of the corresponding cylinder. It is then determined whether the fitted radius and axis are within the preset tolerance range. If they are within the preset tolerance range, coarse alignment is completed. The optimal rigid body transformation is determined by the least squares method. Based on the optimal rigid body transformation, the surface point cloud and the contour edge point cloud are finely registered to unify the surface point cloud and the contour edge point cloud to the same spatial coordinate reference, thus obtaining the registered surface point cloud and the registered contour edge point cloud. Maintain the spatial correspondence between the registered surface point cloud and the registered contour edge point cloud so that surface texture information, contour geometry information and subsequent detection results are all expressed based on a unified spatial coordinate system; Each surface point cloud includes corresponding three-dimensional coordinates and associated original texture color. The contour edge point cloud is used to characterize the actual geometric contour of the bearing roller. The optimal rigid body transformation is the rigid body transformation that minimizes the overall distance between the surface point cloud and the contour edge point cloud in the common geometric feature region.

4. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The process of extracting the characterization features of the unfolded surface of the test surface, determining the surface defect region, recovering the corresponding spatial position through coordinate mapping, and calculating the defect features to obtain the surface defect result includes the following steps: The characterization features of the surface unfolded map to be tested are extracted and similarity calculation is performed with the augmented memory. The cosine distance is used to calculate the nearest neighbor distance between each feature to be tested and the corresponding feature in the augmented memory. The nearest neighbor distance is used as the anomaly score of the corresponding region. Each anomaly score is mapped to the spatial location of the corresponding feature map to obtain anomaly score maps at different scales; each anomaly score map is upsampled to the same size as the original unfolded map, and then fused by taking the maximum value at the pixel level to generate an anomaly heatmap; The Otsu threshold is calculated based on the abnormal heat map, and the segmentation threshold is determined by multiplying the Otsu threshold by a preset magnification factor to generate a binary defect mask. The binarized defect mask is subjected to closing and opening operations in sequence to extract connected components and filter out regions with an area smaller than a preset area threshold to obtain the target defect region. By mapping coordinates, the pixel coordinates corresponding to the target defect area are restored to the world coordinate system and cylindrical coordinate system. The spatial location, geometric features and texture features of the target defect area are calculated. Based on the classifier trained by geometric features, texture features, color features, shape features and context features, the defect type is identified, and the surface defect result is formed.

5. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The process of constructing the spatial association between surface defect regions and contour anomaly regions based on coordinate mapping includes the following steps: Based on the coordinate mapping relationship, the pixel coordinates corresponding to the surface defect area are converted into world coordinates and cylindrical coordinates, and the contour anomaly area and its corresponding world coordinates are determined based on the contour analysis results. The axial distance, circumferential distance, and three-dimensional spatial distance between the surface defect area and the contour abnormal area are calculated respectively. The circumferential distance is calculated based on the difference between the circumferential angular coordinates of the two areas, and the minimum angular distance is taken in combination with the circumferential surrounding characteristics of the cylindrical surface. It is determined whether the axial distance is less than the preset axial tolerance, the circumferential distance is less than the preset circumferential tolerance, and the three-dimensional spatial distance is less than the preset spatial tolerance. If the axial distance, circumferential distance, and three-dimensional spatial distance all meet the corresponding tolerance requirements, it is determined that the surface defect area and the contour abnormal area belong to the same physical area, and a corresponding associated defect pair is constructed. Otherwise, no associated defect pair is constructed. Spatial consistency, positional continuity, and association confidence are calculated for each pair of associated defects. The association confidence is determined by combining spatial consistency, positional continuity, and defect type to construct spatial association relationships. It is also determined whether the surface defect area and the contour anomaly area correspond to the same physical area, which serves as the basis for cross-validation, fusion analysis, and comprehensive quality judgment.

6. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The method includes the following steps: Based on the registered contour edge point cloud, the axis and axis direction are determined, the radial distance corresponding to each contour point is calculated, and the points are sorted according to the axial coordinates to form an ordered contour point set; outlier points are removed by the median absolute deviation filtering method to obtain the contour data after gross error removal. Based on the contour data, a discrete contour curve is constructed, and a smooth spline is used for fitting. An objective function including a fitting error term and a smoothing constraint term is constructed, and the actual contour curve is obtained by optimization. The scaling factor, axial translation, and radial translation between the actual contour curve and the theoretical contour curve are optimized using the least squares method to eliminate the rigid displacement error between the two and obtain the registered actual contour curve. The contour deviation, derivative characteristics, and curvature characteristics between the registered actual contour curve and the theoretical contour curve are calculated, and multi-level contour analysis is performed based on the contour deviation, derivative characteristics, and curvature characteristics. Based on the results of multi-level contour analysis, abnormal contour regions are identified, and the world coordinate positions corresponding to the abnormal contour regions are determined. Contour defect diagnostic information is formed by combining the continuity, change trend and distribution characteristics of the abnormal contour regions, wherein the change characteristics include change trend and distribution characteristics. By combining contour accuracy indicators and contour defect diagnosis information, contour analysis results are generated.

7. The vision-based intelligent detection method for bearing roller surface and contour according to claim 1, characterized in that, The quality assessment using a multi-level decision-making model includes the following steps: A multi-level decision-making model is constructed, which includes a surface defect judgment unit, a contour quality judgment unit, and a fusion judgment unit. The surface quality evaluation result is calculated by the surface defect determination unit based on the defect type, defect area range and defect confidence level. The contour quality evaluation result is calculated by the contour deviation, derivative characteristics, curvature characteristics and contour abnormal area by the contour quality judgment unit. The contour quality judgment adopts the protection zone principle, calculates the expanded uncertainty based on the standard uncertainty and coverage factor of the contour index, and makes a judgment based on the expanded uncertainty correction tolerance zone. Based on the spatial correlation, the fusion judgment unit performs cross-validation and fusion analysis on the surface quality evaluation results and the contour quality evaluation results. If there are correlated defect pairs and the severity combination of the correlated defect pairs exceeds the preset correlation threshold, it is directly judged as unqualified; otherwise, the final quality judgment result is calculated based on the comprehensive weighted calculation of the surface quality evaluation results and the contour quality evaluation results to obtain the final quality judgment result of the bearing roller under test.

8. A vision-based intelligent inspection system for the surface and contour of bearing rollers, characterized in that, The system, configured to perform a vision-based intelligent inspection method for bearing roller surfaces and profiles as described in any one of claims 1 to 7, comprises: A generation module is constructed to build a world coordinate system, simultaneously acquire surface images and two-dimensional contour images of the bearing rollers at the same location, and preprocess them respectively to obtain surface point clouds and contour edge point clouds in the world coordinate system. The alignment and registration module performs fitting, alignment, and registration on the surface point cloud and the contour edge point cloud respectively, to obtain the registered surface point cloud and the registered contour edge point cloud, so as to construct a unified spatial reference. The mapping generation module generates the unfolded map of the surface to be tested and the corresponding coordinate mapping relationship based on the registered surface point cloud, and keeps the surface detection process and the contour detection process using the same spatial coordinate reference. The identification module is determined by constructing an enhanced memory library based on the defect-free surface unfolding atlas, extracting the characterization features of the surface unfolding atlas to be tested, identifying the surface defect region, restoring the corresponding spatial position through coordinate mapping relationship and calculating the defect features to obtain the surface defect result; The module is determined, and the actual contour curve of the bearing roller is reconstructed based on the registered contour edge point cloud. The contour deviation between the actual contour curve and the theoretical contour curve is calculated, the contour anomaly area and its corresponding world coordinates are determined, and the contour analysis results are generated. The fusion analysis module is used to construct the spatial relationship between surface defect areas and contour anomaly areas based on coordinate mapping. The spatial relationship is used to determine whether surface defect areas and contour anomaly areas correspond to the same physical area. Cross-validation and fusion analysis are performed only on surface defect results and contour analysis results that correspond to the same physical area. The comprehensive quality judgment is completed through a multi-level decision model.

Images