Method and system for bearing quality recognition based on machine learning

Abstract

This invention discloses a machine learning-based bearing quality identification method and system. The method includes acquiring the local geometric entropy and weighted deviation field of sample bearing points, identifying defect regions and extracting geometric modal defects based on the weighted deviation field, performing image recognition to obtain visual modal defects, texture features, and color features, calculating texture and color deviations, inputting the raw material characteristics and process parameters of the bearing to be identified into a standard bearing quality prediction model to obtain a standard bearing quality prediction index, inputting the apparent defects and the standard bearing quality prediction index into a bearing quality propagation map to obtain a bearing quality correction index, calculating a quality index, and performing quality identification based on the quality index of the bearing to be identified. This method not only improves the efficiency and accuracy of automotive bearing quality identification but also has good interpretability and can be directly applied to automotive bearing quality identification systems.

Description

Technical Field

[0001] This invention relates to the field of mechanical parts quality inspection technology, and in particular to a bearing quality identification method and system based on machine learning. Background Technology

[0002] As a core supporting component of rotating machinery systems, automotive bearings directly affect the safety and reliability of the entire vehicle. With the automotive industry moving towards lightweighting and high power density, bearing service conditions are becoming increasingly stringent, placing higher demands on the accuracy and comprehensiveness of quality inspection. Traditional bearing quality inspection mainly relies on manual sampling combined with contact measuring tools, which suffers from drawbacks such as low inspection efficiency, strong subjectivity, and easy omissions. In recent years, the introduction of machine vision, laser scanning, and deep learning technologies has driven the development of bearing inspection towards automation and intelligence, enabling rapid identification of surface defects and precise measurement of geometric dimensions.

[0003] However, existing technologies still have significant limitations: First, laser scanning-based geometric detection methods typically use a globally uniform deviation threshold, failing to fully consider the differences in geometric complexity across different functional areas of the bearing, resulting in insufficient sensitivity in identifying shape errors in key mating areas. Simultaneously, visual inspection systems are susceptible to environmental factors such as lighting, temperature, and humidity, and traditional single-factor compensation models struggle to adapt to image quality fluctuations in complex workshop environments. Furthermore, existing quality prediction methods often rely on independent data-driven models based on raw materials and process parameters, lacking a physical correlation mechanism between apparent defects and intrinsic material properties, leading to discrepancies between prediction results and the actual service quality of the bearing. Therefore, this invention proposes a machine learning-based bearing quality identification method and system. By introducing local geometric entropy-weighted deviation field calculation, environmentally coupled image recognition, physical information neural networks, and a quality propagation graph model, it achieves a full-chain quality evaluation from apparent defect detection to service performance prediction, significantly improving the accuracy and reliability of bearing quality identification. This has significant engineering application value for ensuring the safe operation of automotive transmission systems. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing quality identification method and system based on machine learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Laser scanning is used to obtain point clouds of automotive bearing samples. A weighted deviation field is calculated based on the local geometric entropy of the bearing samples. Defect regions are identified and geometric modal defects are extracted based on the weighted deviation field. The image of the car bearing is corrected according to the shooting environment. Image recognition is performed to obtain visual modal defects, texture features and color features, and texture deviation and color deviation are calculated. A standard bearing quality prediction model is constructed based on historical raw material characteristics, historical process parameters, and actual automotive bearing quality indicators. The raw material characteristics and process parameters of the automotive bearing to be identified are input into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. Construct a bearing quality propagation map, input apparent defects and standard bearing quality prediction indicators into the bearing quality propagation map, and obtain bearing quality correction indicators based on physical information for quality propagation; The quality index is calculated based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and the quality is identified based on the quality index of the automotive bearing to be identified. The geometric modal defects include defect location, defect volume, defect area, defect proportion, defect size, shape error, and roundness error; The shooting environment includes light intensity, noise level, temperature, and humidity; The image recognition includes defect detection, feature extraction, and color analysis; The visual modal defects include defect location, defect area, defect proportion, and shape error; The characteristics of the raw materials include the chemical composition, Poisson's ratio, and grain size grade of the raw materials corresponding to each structure of the automotive bearing. The process parameters include the initial forging temperature, heat treatment process curve, and grinding parameters for each structure of the automotive bearing; the grinding parameters include grinding depth, feed rate, and grinding wheel speed. The bearing quality indicators include elastic modulus, hardness, contact fatigue life, fracture toughness, and residual stress. The apparent defects include geometric modal defects, visual modal defects, texture deviations, and color deviations.

[0006] Furthermore, the method for extracting geometric modal defects includes: To identify the automotive bearing, laser scanning is used to acquire the bearing point cloud. The local covariance is calculated, and eigenvalue decomposition is performed to obtain eigenvalues. Based on these eigenvalues, the local geometric entropy is calculated, expressed as: ; ; ; in For point The covariance matrix, For point neighborhood points, For the neighborhood, The number of neighboring points. For the field The local centroid, The matrix of orthogonal eigenvectors It is an eigenvalue diagonal matrix. , , are eigenvalues, and , For point Local geometric entropy; The point cloud of the sample bearing and the point cloud of the standard CAD model are registered using iterative nearest-point registration. The deviation of each point is calculated to construct a weighted deviation field, and a dynamic threshold is calculated to identify defect areas. The expression is as follows: ; ; in For point Weighted bias at the point, For point Deviation at that point The geometric sensitivity coefficient, It is the minimum value of geometric entropy. It is the maximum value of geometric entropy. This is a defective area. For dynamic thresholds, , For the mean and standard deviation of the weighted bias field, This is the threshold sensitivity coefficient; Geometric modal defects are calculated based on the deviation of the defect area.

[0007] Furthermore, the method for calculating texture deviation and color deviation includes: To acquire and correct distortion in images of automotive bearings, a multi-factor coupled correction model is established to simultaneously capture environmental conditions and perform environmental correction on the automotive bearing images. The expression is as follows: ; in The feature vector of the corrected car bearing image. This represents the feature vector of the car bearing image after distortion correction. for Environmental indicator weights, for Environmental indicator values, for Environmental benchmark values This is the temperature drift compensation vector. Temperature deviation; The distortion-corrected and environmentally corrected automotive bearing image is input into a defect detection network based on a pre-trained model to obtain visual modal defects. The gray-level co-occurrence matrix method is used to extract texture features and compare them with the texture features of the standard bearing to calculate the texture deviation. The automotive bearing image is converted to the HSV color space to extract color features and compare them with the color features of the standard bearing to calculate the color deviation.

[0008] Furthermore, the method for obtaining the standard bearing quality prediction index includes: A comprehensive set of bearing quality indicators, raw material characteristics, and process parameters corresponding to historically sampled automotive bearings is collected. The comprehensive set is then randomly divided into a training set and a test set in a 7:3 ratio. The training set is used to train a standard bearing quality prediction model, and the test set is used to test the performance of the standard bearing quality prediction model. The standard bearing quality prediction model includes an input layer, a shared feature extraction layer, a task-specific branch layer, and an output layer. The shared feature extraction layer learns the common influence of different process paths on the microstructure of materials through a fully connected hidden layer, and extracts the shared features between raw materials and process parameters. The task-specific branch layer contains 5 independent branch networks, and obtains bearing quality prediction indicators by regression based on the shared features. The standard bearing quality prediction model is based on a physical information neural network, which combines mean square error loss and mechanical consistency physical constraints to improve the accuracy of the model's predictions. The expression is as follows: ; in For a mixed loss function, For mean square error loss, This represents the batch sample size. Let Hertz contact deformation function represent Sample prediction of contact deformation, for Sample predicted elastic modulus, For the rated normal load, Poisson's ratio, Let be the equivalent radius of curvature of the rolling element and the raceway. for Measured contact deformation of the sample; The raw material characteristics and process parameters of the automotive bearing to be identified are input into the standard bearing quality prediction model to obtain the standard bearing quality prediction index.

[0009] Furthermore, the method for obtaining the bearing quality correction index includes: According to the structural components of the automotive bearing, set up appearance diagram nodes and mass diagram nodes respectively, determine the edge relationships based on the contact relationships of each structural component of the automotive bearing, and construct the bearing mass propagation diagram; The automotive bearing structural components include an inner ring, an outer ring, balls, and a cage; the node features of the appearance map nodes include geometric modal defects, visual modal defects, texture deviations, color deviations, and structural component weights; the node features of the quality map nodes include bearing quality indicators. The edge relationships include internal edge relationships and cross-component edge relationships; the cross-component edge relationships specifically include rolling contact, sliding contact, and potential contact; the edge characteristics of the internal edge relationships include geometric sensitivity coefficient, surface defect sensitivity coefficient, texture deviation weight, and color deviation weight; the edge characteristics of the cross-component edge relationships include contact stiffness, contact stress ratio, relative sliding speed, load distribution coefficient, and fit state; the edge characteristics are determined through bearing mechanics priors. Set the node features of all apparent graph nodes to zero, and set the node features of the quality graph nodes according to the standard bearing quality prediction index to obtain the initial bearing quality propagation graph. Update the node features of all apparent graph nodes using apparent defects, and perform quality propagation to update the initial bearing quality propagation graph to obtain the bearing quality correction index, the expression of which is: ; ; in For graph nodes In the The node features updated in the next iteration. For state update functions, For graph nodes The set of neighboring nodes, For neighboring nodes Node characteristics, For structural sensitivity factor, The structural sensitivity coefficient, For graph nodes with neighboring nodes The mating intervals of corresponding structural components. To design tolerances.

[0010] Furthermore, the method for calculating the quality index includes: The geometric modal defects and visual modal defects are weighted and fused to obtain the fused defects. The geometric deviation rate is calculated based on the fused defects. The weighted geometric deviation rate is calculated according to the structural components. The texture deviation rate and color deviation rate are calculated based on the texture deviation and color deviation. The geometric deviation rate includes volume deviation rate, area deviation rate, size deviation rate, shape deviation rate and roundness deviation rate. The quality index of each structural component is calculated using the weighted geometric deviation rate, texture deviation rate, color deviation rate, and bearing quality correction index. The expression is as follows: ; ; in This refers to the quality index of automotive bearing structural components. , , , For quality weight, This index provides quality indicators, including elastic modulus, hardness, and contact fatigue life. for Correction values ​​for quality indicators for Standard values ​​of quality indicators for Sensitivity coefficient of quality indicators The index includes deviation rates, such as weighted geometric deviation rate, texture deviation rate, and color deviation rate. for Class deviation rate This is a correction value for fracture toughness. This is the required value for fracture toughness. For surface defect indication function, The residual stress index, For residual stress, For ideal residual compressive stress, This represents the tensile stress tolerance threshold. The quality of each structural component is graded according to the quality index. When the quality grade of any structural component does not meet the production requirements, the quality of the automotive bearing is deemed unqualified, and quality traceability is carried out according to the quality grade of each structural component.

[0011] Secondly, an AI-based product quality risk assessment system includes: Point cloud processing module: used to perform laser scanning on automotive bearings to obtain point clouds of the bearing samples, calculate the weighted deviation field based on the local geometric entropy of the bearing samples, identify defect regions and extract geometric modal defects based on the weighted deviation field; Image recognition module: used to correct the car bearing image according to the shooting environment, perform image recognition to obtain visual modal defects, texture features and color features, and calculate texture deviation and color deviation; Quality prediction module: Used to build a standard bearing quality prediction model based on historical raw material characteristics, historical process parameters and actual automotive bearing quality indicators. Input the raw material characteristics and process parameters of the automotive bearing to be identified into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. Quality Correction Module: Used to construct a bearing quality propagation map, inputting apparent defects and standard bearing quality prediction indicators into the bearing quality propagation map, and obtaining bearing quality correction indicators based on physical information for quality propagation; Quality identification module: used to calculate the quality index based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and to perform quality identification based on the quality index of the automotive bearing to be identified.

[0012] The beneficial effects of this invention are: This invention relates to a bearing quality identification method and system based on machine learning. Compared with existing technologies, this invention has the following technical advantages: This invention introduces a local geometric entropy calculation weighted deviation field, which assigns higher weight to key geometric features with drastic curvature changes, thereby more accurately identifying geometric modal defects that affect bearing service performance and significantly improving the accuracy of defect location and quantification. This invention introduces an image correction mechanism based on multiple environmental factors, which eliminates the influence of environmental interference on subsequent image recognition, making the extracted visual modal defects, texture features and color features highly consistent and repeatable. This invention constructs a standard bearing quality prediction model, which can obtain an accurate estimate of the bearing's intrinsic performance based on manufacturing process data before the bearing undergoes destructive testing or long-term fatigue testing, providing a key baseline for subsequent comprehensive quality evaluation. This invention constructs a bearing quality propagation map, which can simulate the propagation chain of various physical deviations. Through quality propagation, a bearing quality correction index that takes into account the system coupling effect is obtained, realizing end-to-end prediction from component-level measurement to system-level performance. This significantly improves the accuracy, robustness, and physical interpretability of automotive bearing quality identification, and has important value for promoting the intelligent and high-quality development of bearing manufacturing. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the steps of the bearing quality identification method based on machine learning according to the present invention. Detailed Implementation

[0014] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0015] The bearing quality identification method and system based on machine learning of this invention includes the following steps: like Figure 1 As shown, this embodiment includes the following steps: Laser scanning is used to obtain point clouds of automotive bearing samples. A weighted deviation field is calculated based on the local geometric entropy of the bearing samples. Defect regions are identified and geometric modal defects are extracted based on the weighted deviation field. The image of the car bearing is corrected according to the shooting environment. Image recognition is performed to obtain visual modal defects, texture features and color features, and texture deviation and color deviation are calculated. A standard bearing quality prediction model is constructed based on historical raw material characteristics, historical process parameters, and actual automotive bearing quality indicators. The raw material characteristics and process parameters of the automotive bearing to be identified are input into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. Construct a bearing quality propagation map, input apparent defects and standard bearing quality prediction indicators into the bearing quality propagation map, and obtain bearing quality correction indicators based on physical information for quality propagation; The quality index is calculated based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and the quality is identified based on the quality index of the automotive bearing to be identified. The geometric modal defects include defect location, defect volume, defect area, defect proportion, defect size, shape error, and roundness error; The shooting environment includes light intensity, noise level, temperature, and humidity; The image recognition includes defect detection, feature extraction, and color analysis; The visual modal defects include defect location, defect area, defect proportion, and shape error; The characteristics of the raw materials include the chemical composition, Poisson's ratio, and grain size grade of the raw materials corresponding to each structure of the automotive bearing. The process parameters include the initial forging temperature, heat treatment process curve, and grinding parameters for each structure of the automotive bearing; the grinding parameters include grinding depth, feed rate, and grinding wheel speed. The bearing quality indicators include elastic modulus, hardness, contact fatigue life, fracture toughness, and residual stress. The apparent defects include geometric modal defects, visual modal defects, texture deviations, and color deviations.

[0016] In this embodiment, the method for extracting geometric modal defects includes: To identify the automotive bearing, laser scanning is used to acquire the bearing point cloud. The local covariance is calculated, and eigenvalue decomposition is performed to obtain eigenvalues. Based on these eigenvalues, the local geometric entropy is calculated, expressed as: ; ; ; in For point The covariance matrix, For point neighborhood points, For the neighborhood, The number of neighboring points. For the field The local centroid, The matrix of orthogonal eigenvectors It is an eigenvalue diagonal matrix. , , are eigenvalues, and , For point Local geometric entropy; The point cloud of the sample bearing and the point cloud of the standard CAD model are registered using iterative nearest-point registration. The deviation of each point is calculated to construct a weighted deviation field, and a dynamic threshold is calculated to identify defect areas. The expression is as follows: ; ; in For point Weighted bias at the point, For point Deviation at that point The geometric sensitivity coefficient, It is the minimum value of geometric entropy. It is the maximum value of geometric entropy. This is a defective area. For dynamic thresholds, , For the mean and standard deviation of the weighted bias field, This is the threshold sensitivity coefficient; Calculate the geometric modal defect based on the defect region deviation; In actual assessment, points covariance matrix It is a 3×3 real matrix, where the orthogonal eigenvector matrix is... The eigenvalues ​​represent the principal directions of the point cloud in a local area. The three eigenvalues ​​of the diagonal matrix represent the variance (dispersion) of the point cloud in the three orthogonal principal directions. Represents the direction of maximum extension (e.g., the axial direction of the raceway). Represents a secondary direction (such as the circumferential direction of the raceway). Represents the direction of minimum variance (usually approximating the normal); when At this time, local isotropy (spherical / chaotic point cloud), local geometric entropy is maximum, and the bearing corresponding region is chamfered / burred / noise; when At that time, local anisotropy is moderate – planar (plate-like) and local geometric entropy is moderate; the corresponding region of the bearing is the end face and the bottom of the oil groove. At this time, the local anisotropy is linear (high curvature) and the local geometric entropy is minimal, and the corresponding region of the bearing is the raceway and the inner ring edge; The calculation points are obtained by iterative registration of nearest points. Deviation at point , To obtain the nearest sample points for the standard CAD model, the geometric sensitivity coefficient is calculated. 0.8 is the threshold sensitivity coefficient. For defect area identification with a value of 0.5, taking points A1, A2, and A3 as examples, the local covariance of the three points is calculated and eigenvalue decomposition is performed to obtain eigenvalues ​​of 0.5 / 0.45 / 0.35, 0.8 / 0.7 / 0.1, and 0.96 / 0.02 / 0.02. The local geometric entropy of the three points is calculated as 1.08 (the geometry of this area is highly complex, corresponding to the chamfer burr area), 0.75 (the geometry of this area is moderately complex, corresponding to the end face oil groove), and 0.15 (the geometry of this area is regular and the curvature is uniform, corresponding to the inner raceway). Based on the global geometric entropy of the automotive bearing to be identified, the dynamic threshold is calculated to be 0.095mm, the minimum geometric entropy value is 0.15, and the maximum geometric entropy value is 1.08. The weighted deviation fields of the three points are calculated to be 0.108mm, 0.137mm, and 0.05mm. Defect identification is performed, and points A1 and A2 are determined to belong to the defect area. Defects are identified by performing defect identification on all scanning points of the automotive bearing to be identified. Based on the standard deviation corresponding to the registration results, the defect location, defect volume, defect area, defect percentage, defect size, shape error, and roundness error of the defect area are calculated to form a geometric modal defect.

[0017] In this embodiment, the method for calculating texture deviation and color deviation includes: To acquire and correct distortion in images of automotive bearings, a multi-factor coupled correction model is established to simultaneously capture environmental conditions and perform environmental correction on the automotive bearing images. The expression is as follows: ; in The feature vector of the corrected car bearing image. This represents the feature vector of the car bearing image after distortion correction. for Environmental indicator weights, for Environmental indicator values, for Environmental benchmark values This is the temperature drift compensation vector. Temperature deviation; The distortion-corrected and environmentally corrected automotive bearing image is input into a defect detection network based on a pre-trained model to obtain visual modal defects. The gray-level co-occurrence matrix method is used to extract texture features and compare them with the texture features of the standard bearing to calculate the texture deviation. The automotive bearing image is converted to the HSV color space to extract color features and compare them with the color features of the standard bearing to calculate the color deviation. In actual evaluation, before acquiring images of automotive bearings, the intrinsic parameter matrix K and distortion coefficient vector (including radial and tangential distortion coefficients) of the industrial camera are first obtained based on Zhang's calibration method. The original automotive bearing images are then distorted according to the pinhole camera model and Brown distortion model (by resampling via bilinear interpolation, automotive bearing images with radial and tangential distortion eliminated are obtained, ensuring the geometric fidelity of circular features such as bearing raceways and cages, providing a basis for subsequent dimensional accuracy measurements). The feature vector of the corrected automotive bearing image includes grayscale values, gradient magnitude, etc. The standard environmental reference values ​​are illuminance of 500 lx, signal-to-noise ratio of 40 dB, temperature of 25 °C, and humidity of 60%. Environmental correction is performed on the automotive bearing image (eliminating shadows caused by uneven lighting, false defects caused by noise, focal length drift caused by temperature and humidity changes, and differences in surface oxidation reflection) to obtain an environmentally normalized automotive bearing image. The distortion-corrected and environmentally corrected images of automobile bearings are input into a defect detection network based on a pre-trained model (using VGG16 pre-trained on ImageNet as the backbone network and YOLOv5 model fine-tuned with bearing defect samples) to obtain visual modal defects. The method for extracting texture features using the gray-level co-occurrence matrix (GLCM) is as follows: The car bearing image is converted into a grayscale image, and a GLCM is constructed at a distance of 1 pixel in four directions: 0°, 45°, 90°, and 135°. Second-order statistical features are calculated: contrast (reflecting the depth of texture grooves), correlation (reflecting the degree of linear dependence of gray levels), energy (reflecting the coarseness of texture), and homogeneity (reflecting local changes in texture). The texture features are then compared with the texture features of a standard bearing to calculate the texture deviation. The automotive bearing image is converted to the HSV color space (hue, saturation, and brightness) to extract color features: color histogram (representing hue distribution), primary hue, secondary hue, average saturation, and brightness (representing surface oxidation and reflectivity). The color features are compared with the color features of the standard bearing to calculate the color deviation (calculated separately for each structure of the automotive bearing). Among them, the hue component deviation is calculated using the annular distance (considering the periodicity of the hue circle), the saturation / brightness deviation is calculated using the absolute difference, and the histogram deviation is calculated using the Bartlett distance.

[0018] In this embodiment, the method for obtaining standard bearing quality prediction indicators includes: A comprehensive set of bearing quality indicators, raw material characteristics, and process parameters corresponding to historically sampled automotive bearings is collected. The comprehensive set is then randomly divided into a training set and a test set in a 7:3 ratio. The training set is used to train a standard bearing quality prediction model, and the test set is used to test the performance of the standard bearing quality prediction model. The standard bearing quality prediction model includes an input layer, a shared feature extraction layer, a task-specific branch layer, and an output layer. The shared feature extraction layer learns the common influence of different process paths on the microstructure of materials through a fully connected hidden layer, and extracts the shared features between raw materials and process parameters. The task-specific branch layer contains 5 independent branch networks, and obtains bearing quality prediction indicators by regression based on the shared features. The standard bearing quality prediction model is based on a physical information neural network, which combines mean square error loss and mechanical consistency physical constraints to improve the accuracy of the model's predictions. The expression is as follows: ; in For a mixed loss function, For mean square error loss, This represents the batch sample size. Let Hertz contact deformation function represent Sample prediction of contact deformation, for Sample predicted elastic modulus, For the rated normal load, Poisson's ratio, Let be the equivalent radius of curvature of the rolling element and the raceway. for Measured contact deformation of the sample; The raw material characteristics and process parameters of the automotive bearing to be identified are input into the standard bearing quality prediction model to obtain the standard bearing quality prediction index. In practical evaluation, the standard bearing quality prediction model uses a shared feature extraction layer with three fully connected hidden layers (dimensions 128, 64, and 32 respectively). ReLU activation and Batch Normalization are employed to extract shared features between raw materials and process parameters. The task-specific branch layer contains five independent branch networks (elastic modulus branch, contact fatigue life branch, fracture toughness branch, hardness branch, and residual stress branch). In the hybrid loss function, the measured contact deformation of the samples is used as the basis for the analysis. Actual contact deformation measurements obtained through indentation experiments or finite element simulations ensure that the elastic modulus predicted by the mechanical consistency physical constraints conforms to Hertz's contact mechanics laws, avoiding non-physical predictions caused by pure data fitting (such as predicting the elastic modulus outside the reasonable range for steel). The equivalent elastic modulus is used in actual calculations; Taking the inner ring of the automotive bearing to be identified (model: DAC35650037 wheel hub bearing unit, used for the front wheels of passenger cars) as an example: The inner ring is made of high carbon chromium bearing steel (grade: GCr15, equivalent to ASTM 52100), and the raw material is hot-rolled annealed bar stock (chemical composition: C-0.98%, Cr-1.52%, Mn-0.32%, Si-0.25%, Mo-0.02%, P-0.008%, S-0.005%), with a Poisson's ratio of 0.3 and a grain size grade of ASTM 9.5. The initial forging temperature of the inner ring is 1180℃ (hot die forging process, using a 2500-ton press; the initial forging temperature ensures the material is in the fully austenitic region and avoids overheating); the processing includes quenching / heating temperature 835℃ / holding time 25min (salt bath furnace) / cooling medium is N32 machine oil (oil temperature 40-60℃) / cooling time 8min, tempering / temperature 160℃ / holding time 120min / air cooling to room temperature; grinding parameters include WA60LV grinding wheel - 35m / s grinding wheel line, 0.8m / min workpiece feed, and 0.005mm grinding depth; Inputting the aforementioned 15-dimensional feature vector into the standard bearing quality prediction model yields the following predicted indicators for the inner ring quality of automotive bearings: elastic modulus 210.5 GPa, hardness 62.3, and contact fatigue life 1.25 × 10⁻⁶. 7 Torque, fracture toughness 28.5 MPa·m 1 / 2 The residual stress is -320 MPa. Similarly, the mass parameters of the remaining structure (outer ring, balls, and cage) are predicted.

[0019] In this embodiment, the method for obtaining bearing quality correction indicators includes: According to the structural components of the automotive bearing, set up appearance diagram nodes and mass diagram nodes respectively, determine the edge relationships based on the contact relationships of each structural component of the automotive bearing, and construct the bearing mass propagation diagram; The automotive bearing structural components include an inner ring, an outer ring, balls, and a cage; the node features of the appearance map nodes include geometric modal defects, visual modal defects, texture deviations, color deviations, and structural component weights; the node features of the quality map nodes include bearing quality indicators. The edge relationships include internal edge relationships and cross-component edge relationships; the cross-component edge relationships specifically include rolling contact, sliding contact, and potential contact; the edge characteristics of the internal edge relationships include geometric sensitivity coefficient, surface defect sensitivity coefficient, texture deviation weight, and color deviation weight; the edge characteristics of the cross-component edge relationships include contact stiffness, contact stress ratio, relative sliding speed, load distribution coefficient, and fit state; the edge characteristics are determined through bearing mechanics priors. Set the node features of all apparent graph nodes to zero, and set the node features of the quality graph nodes according to the standard bearing quality prediction index to obtain the initial bearing quality propagation graph. Update the node features of all apparent graph nodes using apparent defects, and perform quality propagation to update the initial bearing quality propagation graph to obtain the bearing quality correction index, the expression of which is: ; ; in For graph nodes In the The node features updated in the next iteration. For state update functions, For graph nodes The set of neighboring nodes, For neighboring nodes Node characteristics, For structural sensitivity factor, The structural sensitivity coefficient, For graph nodes with neighboring nodes The mating intervals of corresponding structural components. For design tolerances; In actual assessment, the bearing quality propagation diagram actually contains 8 diagram nodes: inner ring apparent node, outer ring apparent node, ball apparent node, cage apparent node, inner ring quality node, outer ring quality node, ball quality node, and quality node cage. The internal edges reflect the transmission path of apparent defects within the same structural component to quality performance (apparent defects correct the theoretical quality of the component through material mechanics mechanisms), totaling 4 edges. The internal edge features describe the sensitivity of apparent defects to propagation to quality parameters. Among them, the geometric sensitivity coefficient represents the influence weight of shape error on the effective value of elastic modulus, taken as exp{-roundness error / form and position tolerance}; the surface defect sensitivity coefficient represents the reduction coefficient of surface roughness / cracks on fatigue life, calculated based on LP life theory; the texture deviation weight represents the influence coefficient of processing texture direction on fracture toughness (anisotropic factor, taken as 0.9 along the texture direction and 1.1 perpendicular to the direction); and the color deviation weight represents the indicative weight of oxidation color on residual stress (based on the empirical coefficient of oxidation color-temperature mapping). The cross-component edge reflects the coupling relationship of quality performance between different structures. Only the nodes of the mass diagram are connected, with a total of 5 undirected edges. Among them, the Hertz contact between the inner raceway and the rolling element, and the outer raceway and the rolling element are rolling contacts, the sliding guide between the cage (rolling element pocket) and the rolling element is a sliding contact, and the collision contact between the inner / outer raceway and the cage is a potential contact. The cross-component edge features describe the coupling strength of quality performance between different components. Among them, the contact stiffness represents the contact deformation resistance based on Hertz theory, the contact stress ratio represents the ratio of the actual contact stress to the material yield strength, the relative sliding velocity represents the relative tangential velocity of the contact interface (0 for rolling contact, non-zero for sliding contact), the load distribution coefficient represents the proportion of the total radial load borne by the contact pair, and the fit state represents the fit clearance or interference. During quality propagation, the state update function uses a gated recurrent unit (GRU) to combine the current node feature state of the graph node with the aggregated neighbor messages to generate a new node feature state. Taking the update of the inner ring quality index of the automotive bearing to be identified (model: DAC35650037 wheel hub bearing unit, used for front wheels of passenger cars) as an example: The inner ring quality diagram nodes were set using automotive bearing quality prediction indicators (where the inner ring quality prediction indicators are elastic modulus 210.5 GPa, hardness 62.3, and contact fatigue life 1.25×10). 7 Torque, fracture toughness 28.5 MPa·m 1 / 2 Residual stress -320 MPa), inner ring surface defect data were obtained through laser scanning and image recognition, and the inner ring surface map nodes were updated (where the geometric modal defect is a roundness error of 0.008 mm, and the visual modal defect is a micro-crack with an area of ​​0.15 mm on the raceway surface). 2 / Defect percentage 0.03%, texture deviation is surface grinding texture roughness out of tolerance / ΔT=0.35, color deviation is slight tempering color and light oxidation / ΔC=0.12, inner ring weight is 0.45kg); set the edge features of the internal edge as follows: geometric sensitivity coefficient 0.449 (the closer the roundness error is to the tolerance, the lower the coefficient), surface defect sensitivity coefficient 0.28 (based on LP theory, roughness 0.4μm compared to the benchmark 0.2μm), texture deviation weight 1.05 (perpendicular to the texture direction, toughness reduction coefficient), color deviation weight (slight oxidation indicates the degree of residual stress release); set the edge feature of the cross-part edge as contact stiffness 2.35×10 6 N / mm, contact stress ratio 0.65, relative sliding speed 0m / s (pure rolling contact), load distribution factor 0.167 (double row bearing, 8 balls per row, single ball load ratio), fit condition +0.002mm (fit clearance) Mass propagation (two-round propagation convergence) was performed to update the initial bearing mass propagation diagram and obtain the bearing mass correction index. Among them, the inner ring mass index is elastic modulus of 198.3 GPa, hardness of 61.9, and contact fatigue life of 8.75 × 10⁻⁶. 6 Torque, fracture toughness 25 MPa·m 1 / 2 Residual stress -284MPa.

[0020] In this embodiment, the method for calculating the quality index includes: The geometric modal defects and visual modal defects are weighted and fused to obtain the fused defects. The geometric deviation rate is calculated based on the fused defects. The weighted geometric deviation rate is calculated according to the structural components. The texture deviation rate and color deviation rate are calculated based on the texture deviation and color deviation. The geometric deviation rate includes volume deviation rate, area deviation rate, size deviation rate, shape deviation rate and roundness deviation rate. The quality index of each structural component is calculated using the weighted geometric deviation rate, texture deviation rate, color deviation rate, and bearing quality correction index. The expression is as follows: ; ; in This refers to the quality index of automotive bearing structural components. , , , For quality weight, This index provides quality indicators, including elastic modulus, hardness, and contact fatigue life. for Correction values ​​for quality indicators for Standard values ​​of quality indicators for Sensitivity coefficient of quality indicators The index includes deviation rates, such as weighted geometric deviation rate, texture deviation rate, and color deviation rate. for Class deviation rate This is a correction value for fracture toughness. This is the required value for fracture toughness. For surface defect indication function, The residual stress index, For residual stress, For ideal residual compressive stress, This represents the tensile stress tolerance threshold. The quality of each structural component is graded according to the quality index. When the quality grade of any structural component does not meet the production requirements, the quality of the automotive bearing is deemed unqualified, and quality traceability is carried out according to the quality grade of each structural component. In practical assessments, the surface defect indicator function is set to 0 in the quality index formula when surface cracks / pits are present, and to 1 when there are no defects. The surface defect indicator function directly determines the validity of the fracture toughness term. There is residual compressive stress when Residual tensile stress exists at this time; Taking the calculation of the inner ring quality index of the automotive bearing to be identified (model: DAC35650037 wheel hub bearing unit, used for the front wheel of a passenger car) as an example, the geometric modal defects (defect volume, defect area, defect size, shape error, roundness error) obtained by laser scanning and the visual modal defects (defect area, shape error) obtained by image recognition are directly averaged and weighted and fused. For similar indicators (such as the defect area in geometric modal defects and the defect area in visual modal defects), the arithmetic mean method is used to group them into a single fused defect indicator. Based on the ratio of the fused defect to the standard design size, the inner ring geometric deviation rate (volume deviation rate 0.8%, area deviation rate 0.5%, size deviation rate 1.2%, shape deviation rate 0.6%, and roundness deviation rate 0.4%) is calculated. The above geometric deviation rates are averaged and weighted to obtain a weighted geometric deviation rate of 0.7%. Based on the texture deviation / color deviation and the corresponding standard texture / color parameters, the texture deviation is calculated to be 1.3% and the color deviation rate to be 9%. The standard values ​​for elastic modulus / hardness / contact fatigue life / fracture toughness are taken as 210 GPa, 62 HRC, and 0.7 × 10⁻⁶, respectively. 7 Rotation, 24 MPa·m 1 / 2 The ideal residual compressive / tensile stress tolerance threshold is taken as -400. For bearings with pressures of 50 MPa and 0.3 / 0.25 / 0.2 / 0.2, the mass index of the inner ring is calculated to be 0.851. Similarly, the mass indices of the outer ring, balls, and cage are calculated to be 0.785 / 0.92 / 0.658. Based on the mass indices, each structural component is graded (Grade A: BQCI ≥ 0.90, minimal geometric deviation, excellent physical properties, and sufficient residual compressive stress; Grade B: 0.70 ≤ BQCI < 0.90, minor defects or performance deviations, meeting normal service requirements; Grade C: BQCI < 0.70, severe defects or significant performance degradation, with a risk of early failure). The mass grades of the inner ring, outer ring, balls, and cage are B / B / A / C. According to production requirements (all structural components must reach Grade B or above), the cage's Grade C does not meet the requirements. Therefore, the automotive bearing is substandard, and the raw materials and production process of the cage are traced back to their source.

[0021] Secondly, a machine learning-based bearing quality identification system includes: Point cloud processing module: used to perform laser scanning on automotive bearings to obtain point clouds of the bearing samples, calculate the weighted deviation field based on the local geometric entropy of the bearing samples, identify defect regions and extract geometric modal defects based on the weighted deviation field; Image recognition module: used to correct the car bearing image according to the shooting environment, perform image recognition to obtain visual modal defects, texture features and color features, and calculate texture deviation and color deviation; Quality prediction module: Used to build a standard bearing quality prediction model based on historical raw material characteristics, historical process parameters and actual automotive bearing quality indicators. Input the raw material characteristics and process parameters of the automotive bearing to be identified into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. Quality Correction Module: Used to construct a bearing quality propagation map, inputting apparent defects and standard bearing quality prediction indicators into the bearing quality propagation map, and obtaining bearing quality correction indicators based on physical information for quality propagation; Quality identification module: used to calculate the quality index based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and to perform quality identification based on the quality index of the automotive bearing to be identified.

[0022] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bearing quality identification method based on machine learning, characterized in that, Includes the following steps: S1. Perform laser scanning on the automotive bearing to obtain the point cloud of the bearing sample. Calculate the weighted deviation field based on the local geometric entropy of the bearing sample. Identify the defect region and extract the geometric modal defect based on the weighted deviation field. S2. Correct the car bearing image according to the shooting environment, perform image recognition to obtain visual modal defects, texture features and color features, and calculate texture deviation and color deviation; S3. Construct a standard bearing quality prediction model based on historical raw material characteristics, historical process parameters and actual automotive bearing quality indicators. Input the raw material characteristics and process parameters of the automotive bearing to be identified into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. S4. Construct a bearing quality propagation diagram, input the apparent defects and standard bearing quality prediction indicators into the bearing quality propagation diagram, and obtain bearing quality correction indicators based on physical information for quality propagation. S5. Calculate the quality index based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and perform quality identification based on the quality index of the automotive bearing to be identified. The geometric modal defects include defect location, defect volume, defect area, defect proportion, defect size, shape error, and roundness error; The shooting environment includes light intensity, noise level, temperature, and humidity; The image recognition includes defect detection, feature extraction, and color analysis; The visual modal defects include defect location, defect area, defect proportion, and shape error; The characteristics of the raw materials include the chemical composition, Poisson's ratio, and grain size grade of the raw materials corresponding to each structure of the automotive bearing. The process parameters include the initial forging temperature, heat treatment process curve, and grinding parameters for each structure of the automotive bearing; the grinding parameters include grinding depth, feed rate, and grinding wheel speed. The bearing quality indicators include elastic modulus, hardness, contact fatigue life, fracture toughness, and residual stress. The apparent defects include geometric modal defects, visual modal defects, texture deviations, and color deviations.

2. The bearing quality identification method based on machine learning according to claim 1, characterized in that, The method for extracting geometric modal defects includes: To identify the automotive bearing, laser scanning is used to acquire the bearing point cloud. The local covariance is calculated, and eigenvalue decomposition is performed to obtain eigenvalues. Based on these eigenvalues, the local geometric entropy is calculated, expressed as: ； ； ； in For point The covariance matrix, For point neighborhood points, For the neighborhood, The number of neighboring points. For the field The local centroid, The matrix of orthogonal eigenvectors It is an eigenvalue diagonal matrix. , , are eigenvalues, and , For point Local geometric entropy; The point cloud of the sample bearing and the point cloud of the standard CAD model are registered using iterative nearest-point registration. The deviation of each point is calculated to construct a weighted deviation field, and a dynamic threshold is calculated to identify defect areas. The expression is as follows: ； ； in For point Weighted bias at the point, For point Deviation at that point The geometric sensitivity coefficient, It is the minimum value of geometric entropy. It is the maximum value of geometric entropy. This is a defective area. For dynamic thresholds, , For the mean and standard deviation of the weighted bias field, This is the threshold sensitivity coefficient; Geometric modal defects are calculated based on the deviation of the defect area.

3. The bearing quality identification method based on machine learning according to claim 1, characterized in that, The method for calculating texture deviation and color deviation includes: To acquire and correct distortion in images of automotive bearings, a multi-factor coupled correction model is established to simultaneously capture environmental conditions and perform environmental correction on the automotive bearing images. The expression is as follows: ； in The feature vector of the corrected car bearing image. This represents the feature vector of the car bearing image after distortion correction. for Environmental indicator weights, for Environmental indicator values, for Environmental benchmark values This is the temperature drift compensation vector. Temperature deviation; The distortion-corrected and environmentally corrected automotive bearing image is input into a defect detection network based on a pre-trained model to obtain visual modal defects. The gray-level co-occurrence matrix method is used to extract texture features and compare them with the texture features of the standard bearing to calculate the texture deviation. The automotive bearing image is converted to the HSV color space to extract color features and compare them with the color features of the standard bearing to calculate the color deviation.

4. The bearing quality identification method based on machine learning according to claim 1, characterized in that, The method for obtaining standard bearing quality prediction indicators includes: A comprehensive set of bearing quality indicators, raw material characteristics, and process parameters corresponding to historically sampled automotive bearings is collected. The comprehensive set is then randomly divided into a training set and a test set in a 7:3 ratio. The training set is used to train a standard bearing quality prediction model, and the test set is used to test the performance of the standard bearing quality prediction model. The standard bearing quality prediction model includes an input layer, a shared feature extraction layer, a task-specific branch layer, and an output layer. The shared feature extraction layer learns the common influence of different process paths on the microstructure of materials through a fully connected hidden layer, and extracts the shared features between raw materials and process parameters. The task-specific branch layer contains 5 independent branch networks, and obtains bearing quality prediction indicators by regression based on the shared features. The standard bearing quality prediction model is based on a physical information neural network, which combines mean square error loss and mechanical consistency physical constraints to improve the accuracy of the model's predictions. The expression is as follows: ； in For a mixed loss function, For mean square error loss, This represents the batch sample size. Let Hertz contact deformation function represent Sample prediction of contact deformation, for Sample predicted elastic modulus, For the rated normal load, Poisson's ratio, Let be the equivalent radius of curvature of the rolling element and the raceway. for Actual contact deformation of the sample; The raw material characteristics and process parameters of the automotive bearing to be identified are input into the standard bearing quality prediction model to obtain the standard bearing quality prediction index.

5. The bearing quality identification method based on machine learning according to claim 1, characterized in that, The method for obtaining bearing quality correction indicators includes: According to the structural components of the automotive bearing, set up appearance diagram nodes and mass diagram nodes respectively, determine the edge relationships based on the contact relationships of each structural component of the automotive bearing, and construct the bearing mass propagation diagram; The automotive bearing structural components include an inner ring, an outer ring, balls, and a cage; the node features of the appearance map nodes include geometric modal defects, visual modal defects, texture deviations, color deviations, and structural component weights; the node features of the quality map nodes include bearing quality indicators. The edge relationships include internal edge relationships and cross-component edge relationships; the cross-component edge relationships specifically include rolling contact, sliding contact, and potential contact; the edge characteristics of the internal edge relationships include geometric sensitivity coefficient, surface defect sensitivity coefficient, texture deviation weight, and color deviation weight; the edge characteristics of the cross-component edge relationships include contact stiffness, contact stress ratio, relative sliding speed, load distribution coefficient, and fit state; the edge characteristics are determined through bearing mechanics priors. Set the node features of all apparent graph nodes to zero, and set the node features of the quality graph nodes according to the standard bearing quality prediction index to obtain the initial bearing quality propagation graph. Update the node features of all apparent graph nodes using apparent defects, and perform quality propagation to update the initial bearing quality propagation graph to obtain the bearing quality correction index, the expression of which is: ； ； in For graph nodes In the The node features updated in the next iteration. For state update functions, For graph nodes The set of neighboring nodes, For neighboring nodes Node characteristics, For structural sensitivity factor, The structural sensitivity coefficient, For graph nodes with neighboring nodes The mating intervals of corresponding structural components. To design tolerances.

6. The bearing quality identification method based on machine learning according to claim 1, characterized in that, The method for calculating the quality index includes: The geometric modal defects and visual modal defects are weighted and fused to obtain the fused defects. The geometric deviation rate is calculated based on the fused defects. The weighted geometric deviation rate is calculated according to the structural components. The texture deviation rate and color deviation rate are calculated based on the texture deviation and color deviation. The geometric deviation rate includes volume deviation rate, area deviation rate, size deviation rate, shape deviation rate and roundness deviation rate. The quality index of each structural component is calculated using the weighted geometric deviation rate, texture deviation rate, color deviation rate, and bearing quality correction index. The expression is as follows: ； ； in This refers to the quality index of automotive bearing structural components. , , , For quality weight, This index provides quality indicators, including elastic modulus, hardness, and contact fatigue life. for Correction values ​​for quality indicators for Standard values ​​of quality indicators for Sensitivity coefficient of quality indicators The index includes deviation rates, such as weighted geometric deviation rate, texture deviation rate, and color deviation rate. for Class deviation rate This is a correction value for fracture toughness. This is the required value for fracture toughness. For surface defect indication function, The residual stress index, For residual stress, For ideal residual compressive stress, This represents the tensile stress tolerance threshold. The quality of each structural component is graded according to the quality index. When the quality grade of any structural component does not meet the production requirements, the quality of the automotive bearing is deemed unqualified, and quality traceability is carried out according to the quality grade of each structural component.

7. A machine learning-based bearing quality identification system for performing the method described in any one of claims 1-6, characterized in that, include: Point cloud processing module: used to perform laser scanning on automotive bearings to obtain point clouds of the bearing samples, calculate the weighted deviation field based on the local geometric entropy of the bearing samples, identify defect regions and extract geometric modal defects based on the weighted deviation field; Image recognition module: used to correct the car bearing image according to the shooting environment, perform image recognition to obtain visual modal defects, texture features and color features, and calculate texture deviation and color deviation; Quality prediction module: Used to build a standard bearing quality prediction model based on historical raw material characteristics, historical process parameters and actual automotive bearing quality indicators. Input the raw material characteristics and process parameters of the automotive bearing to be identified into the standard bearing quality prediction model to obtain the standard bearing quality prediction indicators. Quality Correction Module: Used to construct a bearing quality propagation map, inputting apparent defects and standard bearing quality prediction indicators into the bearing quality propagation map, and obtaining bearing quality correction indicators based on physical information for quality propagation; Quality identification module: used to calculate the quality index based on the apparent defects of the automotive bearing to be identified and the bearing quality correction index, and to perform quality identification based on the quality index of the automotive bearing to be identified.

Images