Tracing, positioning and detecting system for surface defects of automobile manufacturing parts

Abstract

The invention discloses an automobile manufacturing part surface defect tracing and positioning detection system, and belongs to the technical field of industrial defect detection. Comprising an isotropic scattering reference point module, the surface of which is provided with a specific micron-sized three-dimensional geometric structure; the multi-view synchronous acquisition module is deployed in the circumferential direction of an automobile part detection station, comprises at least three industrial cameras which are arranged in an elevation angle difference manner, and acquires a multi-angle original image with a reference point through a synchronous triggering acquisition device; the scattering feature intelligent positioning module is in signal connection with the multi-view synchronous acquisition module; the dynamic calibration compensation module is used for receiving the initial registration result output by the multi-view geometric resolving unit and correcting a scattering compensation coefficient of the reference point in a high-curvature area in real time according to curved surface curvature distribution; and outputting the corrected high-precision three-dimensional reconstruction data. Through a refraction compensation formula, material optical parameters and geometric curvature are fused, and positioning error multiple compression is achieved in high-distortion areas such as vehicle door ridges.

Description

Technical Field

[0001] This invention relates to the field of industrial defect detection technology, and more specifically, to a system for tracing and locating surface defects in automotive manufacturing parts. Background Technology

[0002] In the automated manufacturing and assembly of automotive parts, especially critical exterior components and high-precision structural parts such as doors, hoods, and fenders, online, non-contact visual inspection of their surface quality and geometric dimensions is crucial. This type of inspection often requires the use of multiple industrial cameras to simultaneously or sequentially capture image information of the target area from different angles. This overcomes the problems of localized occlusion or information loss caused by the geometric complexity, high curvature, and highly reflective surface characteristics of the parts themselves from a single viewpoint.

[0003] However, existing multi-view vision-based positioning technologies face a critical bottleneck: the projection imaging of key physical reference points used for camera extrinsic parameter calibration or point cloud stitching on highly reflective, high-curvature automotive parts, such as galvanized steel stampings and high-gloss painted surfaces, is severely hampered by multi-source, uncontrollable specular reflections and high-frequency speckle effects. These reference points are not ideal diffuse reflection point sources, especially under low-angle illumination or multi-camera observation angles. Their imaging edges exhibit significant "halo diffusion" and "position blurring" due to strong specular reflections, severely distorting the originally clear, high-contrast point outline geometric center imaging.

[0004] Existing technological solutions to this problem have significant drawbacks:

[0005] 1. Reference point distortion under reflective curved surfaces: Highly reflective materials such as galvanized steel plates or high-gloss paint surfaces cause halo diffusion and speckle noise in the imaging of reference points. Existing marker points are affected when the curvature is >50m. -1 The center positioning error of the area is as high as 0.35mm, which cannot meet the process requirement of ±0.05mm for the surface difference of the vehicle body gap.

[0006] 2. Systemic failures due to dynamic interference: Temperature differences in the workshop cause thermal deformation of metal parts, and traditional vision systems have no compensation mechanism; stamping oil film covers the reference point, and the attenuation of scattering characteristics leads to a decrease in algorithm accuracy; and sudden changes in ambient light cause image overexposure or underexposure, and HDR fusion is not coupled with feature extraction, resulting in a decrease in positioning consistency. Summary of the Invention

[0007] Based on the above-mentioned deficiencies, the purpose of this invention is to provide a system for tracing and locating surface defects of automotive manufacturing parts; and to solve the systemic failure problems of reference point distortion and dynamic interference under the above-mentioned reflective curved surfaces.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A system for tracing and locating surface defects in automotive manufacturing parts includes:

[0010] An isotropic scattering reference point module has a specific micron-scale three-dimensional geometric structure on its surface, which is used to generate a highly uniform diffuse scattering light field in the incident angle range of 10°-80° and suppress specular reflection.

[0011] The multi-view synchronous acquisition module is deployed around the automotive parts inspection station and includes at least three industrial cameras arranged at different elevation angles. It acquires multi-angle raw images with reference points by synchronously triggering the acquisition device.

[0012] A scattering feature intelligent positioning module, signal-connected to the multi-view synchronous acquisition module, is used to convert physical scattering characteristics into mathematically solvable spatial coordinate data, specifically including:

[0013] Scattering schlieren feature extraction unit: Based on the optical properties of the reference point microstructure, it separates the scattering region from the multi-view original image, calculates the radial brightness gradient distribution, and generates a feature vector;

[0014] Sub-pixel registration unit: Through a pre-trained triple attention convolutional neural network, the feature vector is matched with the reference scattering schlieren template, and the center coordinates of the reference point with an accuracy better than 0.05 pixels are output.

[0015] Multi-view geometry solving unit: Establishes matching point pairs based on the sub-pixel center coordinates, iteratively solves the camera extrinsic parameter matrix using the Zhang Zhengyou calibration method, and performs weighted RANSAC point cloud initial registration;

[0016] The dynamic calibration and compensation module receives the initial registration results output by the multi-view geometry solution unit, corrects the scattering compensation coefficient of the reference point in the high curvature region in real time according to the surface curvature distribution, and outputs the corrected high-precision 3D reconstruction data.

[0017] Furthermore, the isotropic scattering reference point module includes:

[0018] Microcavity array unit: A conical microcavity with a sidewall tilt angle of θ = 55 ± 2° is formed by etching the surface of an alumina ceramic substrate, which is used to generate a uniform scattering field in the range of 10°-80° incident angle;

[0019] Anti-reflective coating unit: A layer of silica nanoparticles coated on the surface of the microcavity, with titanium dioxide photocatalytic particles embedded on its surface to decompose the attached oil film and maintain scattering characteristics;

[0020] Mechanical positioning unit: It has a built-in electromagnet and flexible universal joint, which makes the reference point adaptively fit the curved surface of the workpiece.

[0021] Furthermore, the convolutional neural network is trained through the following steps:

[0022] During the sample collection phase, curvature values ​​of 0-200 μm were covered on the surfaces of galvanized steel sheets, high-gloss polyurethane paint surfaces, and carbon fiber composite materials. -1 In the test area, millions of frames of reference point scattered light spot image samples were collected using multi-angle industrial cameras;

[0023] In the feature engineering stage, a 16-direction radial brightness gradient histogram with the reference point center as the origin is extracted for each frame of scattered light spot image. The average gradient magnitude is calculated by dividing the angle interval into 0°-360° equally and combining them into a 128-dimensional schlieren feature vector.

[0024] During the network optimization phase, a triple attention mechanism is constructed by using cascaded channel attention branches, spatial attention branches, and curvature attention branches to optimize feature matching.

[0025] Furthermore, the sub-pixel registration unit includes:

[0026] Curvature-responsive compensation mechanism: When the received real-time curvature K of the surface is greater than 50m -1 At that time, retrieve the refractive index η of the corresponding material;

[0027] And activate the bilinear interpolation algorithm to calculate the edge refraction distortion compensation amount, the formula is:

[0028] δ=0.02·|K|·(η-1);

[0029] Corrected reference point center coordinates: (x', y') = (x + δ) x ,y+δ y );

[0030] In the formula: (x,y) are the coordinates of the center of the initially detected reference point; δ x and δ y Here are the distortion compensation amounts in the X and Y directions; (x', y') are the corrected real physical coordinates.

[0031] δ represents the edge refraction distortion compensation amount, 0.02 is the curvature-refraction coupling coefficient, determined through ray tracing simulation and fitting with actual vehicle data; |K| is the absolute value of the surface curvature obtained by point cloud computing, with an effective range of 50-250m. -1 η is the relative refractive index of the workpiece, mapped through a pre-stored material database;

[0032] Furthermore, the subpixel registration unit also includes an illumination fluctuation response mechanism: when the ambient illumination sensor detects a fluctuation amplitude >15%, the industrial camera is controlled to switch to HDR mode, and three layers of imaging are performed with three exposure times. The schlieren feature vectors of each exposure layer image are extracted and weighted by the reciprocal of the exposure time to output the subpixel center coordinates that are resistant to glare interference.

[0033] Furthermore, the multi-view geometry solving unit is also used to perform thermal expansion compensation:

[0034] Receive the workpiece surface temperature T collected by the infrared sensor and apply the deformation compensation matrix:

[0035] [Δx,Δy]=[α·(T-T0)·L,β·(T-T0)·L]

[0036] In the formula: α and β are the coefficients of thermal expansion of the workpiece material (α = 11.8 × 10⁻⁶ for galvanized steel). -6 / ℃, β=9.5×10 -6 / ℃), pre-stored in the process database; T0 is the system calibration reference temperature (23℃±1℃, conforming to ISO Class 1 tolerance); L is the Euclidean distance from the reference point to the nearest fixed edge of the workpiece, in mm;

[0037] The initial registration coordinates are corrected to: (x', y') = (x + Δx, y + Δy)

[0038] In the formula: (x,y) are the initial detected coordinates of the reference point center, which have thermal deformation error; Δx and Δy are the compensation displacement caused by thermal expansion, calculated by the deformation matrix; (x',y') are the corrected real physical coordinates, i.e., eliminating temperature error.

[0039] Furthermore, in the dynamic calibration compensation module, the formula for calculating the scattering compensation coefficient in real time based on the workpiece material type and local surface curvature is as follows:

[0040] γ=1+λ×|K|

[0041] In the formula: λ is the dynamic material coefficient, which is determined by mapping through a pre-stored database; |K| is the absolute value of the surface curvature, which is calculated by the difference of the point cloud normal vector; 1 is the baseline, i.e., the plane;

[0042] The compensation process is as follows: the initial registered point cloud coordinates (x0, y0, z0) are corrected to (x, y, z) = γ·(x0, y0, z0), where (x, y, z) are the corrected true coordinates.

[0043] Furthermore, it also includes a closed-loop error convergence module, connected to the output of the dynamic calibration compensation module. The closed-loop error convergence module includes:

[0044] The benchmark comparison unit is used to perform ICP matching between the corrected 3D reconstructed point cloud and the laser tracker benchmark point cloud to calculate the Hausdorff distance ε.

[0045] The dynamic optimization unit for the λ value updates the material coefficient λ of weight 7 in reverse iteration when ε > 0.03 mm.

[0046] λ new =λ old +sign(Δε)×0.0005

[0047] In the formula: λ old λ is the material coefficient used in the current batch. new The material coefficient for the next batch of optimization; Δε = current error - previous error, sign is the sign function;

[0048] The historical error database construction unit stores the optimal λ value under different material and curvature combinations, forming a process knowledge graph;

[0049] The convergence verification unit locks the current λ value and outputs a calibration completion signal when ε≤0.02mm is reached for 3 consecutive batches.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] 1. When in use, this automotive manufacturing parts surface defect traceability and positioning detection system uses a 55° tilted microcavity array design to accurately suppress low-angle specular reflection halo, achieving an 82% reduction in reflection intensity under an 80° incident angle condition; combined with an anti-reflective coating to decompose the attached oil film, the positioning error in oily environments is compressed from 3.2 pixels to 0.4 pixels; and with the help of a mechanical positioning unit, the reference point adaptively fits the workpiece surface.

[0052] 2. When using this automotive manufacturing parts surface defect traceability and positioning detection system, it integrates material optical parameters and geometric curvature through a refraction compensation formula to achieve a multiple compression of positioning error in high distortion areas such as door edges; it develops a triple attention convolutional network to enhance feature extraction in high curvature areas, thereby improving the detection accuracy of carbon fiber composites; and it constructs a dynamic scattering compensation mechanism to solve point cloud distortion caused by material differences, pushing the cross-material detection consistency of galvanized steel and aluminum alloys to over 98%.

[0053] 3. When in use, this automotive manufacturing parts surface defect traceability and positioning detection system uses a thermal deformation compensation matrix to offset the thermal expansion error of metal parts in real time; it adopts an HDR three-exposure reciprocal weighted fusion strategy to couple feature extraction and illumination adaptation, thereby improving the positioning stability in strong light fluctuation environments; it pioneers a coefficient self-optimization engine to cope with material batch fluctuations, driving a further improvement in detection accuracy in the first month of mass production, and achieves continuous evolution through multiple sets of process knowledge graphs. Attached Figure Description

[0054] Figure 1 This is an overall flowchart of a surface defect tracing and location detection system for automotive manufacturing parts.

[0055] Figure 2 This is an overall diagram of the intelligent positioning module based on scattering characteristics in a traceability and location detection system for surface defects in automotive manufacturing parts.

[0056] Figure 3 This is an overall diagram of the isotropic scattering reference point module of a surface defect tracing and location detection system for automotive manufacturing parts. Detailed Implementation

[0057] Please see Figure 1 - Figure 3 A system for tracing and locating surface defects in automotive manufacturing parts includes: an isotropic scattering reference point module, the surface of which is constructed with a specific micron-level three-dimensional geometric structure to generate a highly uniform diffuse scattering light field within the incident angle range of 10°-80° and suppress specular reflection.

[0058] The multi-view synchronous acquisition module is deployed around the automotive parts inspection station and includes at least three industrial cameras arranged at different elevation angles. It acquires multi-angle raw images with reference points by synchronously triggering the acquisition device.

[0059] A scattering feature intelligent positioning module, signal-connected to the multi-view synchronous acquisition module, is used to convert physical scattering characteristics into mathematically solvable spatial coordinate data, specifically including:

[0060] Scattering schlieren feature extraction unit: Based on the optical properties of the reference point microstructure, it separates the scattering region from the multi-view original image, calculates the radial brightness gradient distribution, and generates a feature vector;

[0061] Sub-pixel registration unit: Through a pre-trained triple attention convolutional neural network, the feature vector is matched with the reference scattering schlieren template, and the center coordinates of the reference point with an accuracy better than 0.05 pixels are output.

[0062] Multi-view geometry solving unit: Establishes matching point pairs based on the sub-pixel center coordinates, iteratively solves the camera extrinsic parameter matrix using the Zhang Zhengyou calibration method, and performs weighted RANSAC point cloud initial registration;

[0063] The dynamic calibration and compensation module receives the initial registration results output by the multi-view geometry solution unit, corrects the scattering compensation coefficient of the reference point in the high curvature region in real time according to the surface curvature distribution, and outputs the corrected high-precision 3D reconstruction data.

[0064] In a preferred embodiment, the isotropic scattering reference point module includes:

[0065] Microcavity array unit: A conical microcavity with a diameter of 40-80μm, an aspect ratio of 0.5-1.2, and a sidewall tilt angle of θ=55±2° is formed by etching the surface of an alumina ceramic substrate. It is used to generate a uniform scattering field in the incident angle range of 10°-80° and solve the problem of specular reflection halo when observing at low angles.

[0066] Anti-reflective coating unit: A layer of silica nanoparticles coated on the surface of the microcavity, with titanium dioxide photocatalytic particles embedded on its surface, used to decompose the attached oil film and maintain scattering characteristics; solving the scattering failure caused by stamping oil film coverage;

[0067] Mechanical positioning unit: It has a built-in electromagnet and flexible universal joint, which makes the reference point adaptively fit the curved surface of the workpiece.

[0068] In this preferred embodiment, the convolutional neural network is trained through the following steps:

[0069] During the sample collection phase, curvature values ​​of 0-200 μm were covered on the surfaces of galvanized steel sheets, high-gloss polyurethane paint surfaces, and carbon fiber composite materials. -1 In the test area, millions of frames of reference point scattered light spot image samples were collected using multi-angle industrial cameras;

[0070] In the feature engineering stage, a 16-direction radial brightness gradient histogram with the reference point center as the origin is extracted for each frame of scattered light spot image. The average gradient magnitude is calculated by dividing the angle interval into 0°-360° equally and combining them into a 128-dimensional schlieren feature vector.

[0071] During the network optimization phase, a triple attention mechanism is constructed by using cascaded channel attention branches, spatial attention branches, and curvature attention branches to optimize feature matching.

[0072] In this preferred embodiment, the sub-pixel registration unit includes:

[0073] Curvature-responsive compensation mechanism: When the received real-time curvature K of the surface is greater than 50m -1 When retrieving the corresponding material refractive index η, in this embodiment, for example, η = 2.42 for galvanized steel plate and η = 1.55 for high-gloss paint.

[0074] And activate the bilinear interpolation algorithm to calculate the edge refraction distortion compensation amount, the formula is:

[0075] δ=0.02·|K|·(η-1);

[0076] Corrected reference point center coordinates: (x', y') = (x + δ) x ,y+δ y );

[0077] In the formula: (x,y) are the coordinates of the center of the initially detected reference point; δx and δ y Here are the distortion compensation amounts in the X and Y directions; (x', y') are the corrected real physical coordinates.

[0078] δ represents the edge refraction distortion compensation amount, 0.02 is the curvature-refraction coupling coefficient, determined through ray tracing simulation and fitting with actual vehicle data; |K| is the absolute value of the surface curvature obtained by point cloud computing, with an effective range of 50-250m. -1 η is the relative refractive index of the workpiece, mapped through a pre-stored material database: galvanized steel sheet η = 2.42, high-gloss polyurethane paint η = 1.55;

[0079] Curvature > 50m -1 In areas such as the edges of car doors, light refraction through highly reflective curved surfaces can cause a shift in the image position of the reference point. The shift δ is strongly correlated with the curvature K and the refractive index η of the material. The theoretical shift δ is calculated using a physical model, and the coordinates are pulled back to the true position using additive inverse compensation.

[0080] In a preferred embodiment, the subpixel registration unit further includes an illumination fluctuation response mechanism: when the ambient illumination sensor detects a fluctuation amplitude >15%, the industrial camera is controlled to switch to HDR mode, and three layers of imaging are performed with three exposure times. The schlieren feature vectors of each exposure layer image are extracted and fused by weighting the inverse of the exposure time to output the subpixel center coordinates that are resistant to glare interference.

[0081] In this preferred embodiment, the multi-view geometry solving unit is also used to perform thermal expansion compensation:

[0082] Receive the workpiece surface temperature T collected by the infrared sensor and apply the deformation compensation matrix:

[0083] [Δx,Δy]=[α·(T-T0)·L,β·(T-T0)·L]

[0084] In the formula: α and β are the coefficients of thermal expansion of the workpiece material (α = 11.8 × 10⁻⁶ for galvanized steel). -6 / ℃, β=9.5×10 -6 / ℃), pre-stored in the process database; T0 is the system calibration reference temperature (23℃±1℃, conforming to ISO Class 1 tolerance); L is the Euclidean distance from the reference point to the nearest fixed edge of the workpiece, in mm;

[0085] The initial registration coordinates are corrected to: (x', y') = (x + Δx, y + Δy)

[0086] In the formula: (x,y) are the initial detected coordinates of the reference point center, which have thermal deformation error; Δx and Δy are the compensation displacement caused by thermal expansion, calculated by the deformation matrix; (x',y') are the corrected true physical coordinates, i.e., eliminating temperature error.

[0087] Temperature differences on the automotive production line cause metal parts to expand and contract. For example, a car body will shrink by 0.1 mm / m when it moves from the paint booth (60℃) to the final assembly line (25℃). By acquiring the workpiece temperature T in real time through infrared thermography and utilizing pre-stored material parameters α and β (anisotropic expansion rate of galvanized steel), and based on the leverage effect (the larger the L value, the greater the compensation), temperature drift compensation is ultimately achieved at the sub-pixel level.

[0088] In this preferred embodiment, the formula for calculating the scattering compensation coefficient in the dynamic calibration compensation module based on the workpiece material type and local surface curvature is as follows:

[0089] γ=1+λ×|K|

[0090] In the formula: λ is the dynamic material coefficient, which is determined through mapping from a pre-stored database, for example,

[0091] λ = 0.002 (carbon fiber composite, low reflectivity with weak compensation); λ = 0.005 (aluminum alloy, normal reflectivity with compensation); λ = 0.007 (galvanized steel sheet, high reflectivity requires strong compensation);

[0092] |K| is the absolute value of the surface curvature, which is calculated by the difference of the point cloud normal vectors; 1 is the baseline, i.e., the plane;

[0093] The compensation is performed as follows: the initial registered point cloud coordinates (x0, y0, z0) are corrected to (x, y, z) = γ·(x0, y0, z0).

[0094] The formula uses the absolute value |K| to address the asymmetry of scattering distortion between convex and concave surfaces; λ, as a material coefficient, is not a fixed value but is dynamically updated through transfer learning; the compensation mechanism is directly linked to the subsequent point cloud reconstruction.

[0095] In a preferred embodiment, the module further includes a closed-loop error convergence module connected to the output of the dynamic calibration compensation module. The closed-loop error convergence module includes:

[0096] The benchmark comparison unit is used to perform ICP matching between the corrected 3D reconstructed point cloud and the laser tracker benchmark point cloud to calculate the Hausdorff distance ε.

[0097] The dynamic optimization unit for the λ value updates the material coefficient λ of weight 7 in reverse iteration when ε > 0.03 mm.

[0098] λ new =λ old+sign(Δε)×0.0005

[0099] In the formula: λ old λ is the material coefficient used in the current batch. new The material coefficient for the next batch of optimization; Δε = current error - previous error, sign is the sign function;

[0100] The historical error database construction unit stores the optimal λ value under different material and curvature combinations, forming a process knowledge graph;

[0101] The convergence verification unit locks the current λ value and outputs a calibration completion signal when ε≤0.02mm is reached for 3 consecutive batches.

[0102] Due to the differences in reflectivity of galvanized sheets from different steel mills, a closed-loop error convergence module is used to address the λ value drift caused by batch fluctuations in materials, forming a self-learning process library that continuously improves compensation accuracy over mass production time.

[0103] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for tracing and locating surface defects in automotive manufacturing parts, characterized in that: include: An isotropic scattering reference point module has a specific micron-scale three-dimensional geometric structure on its surface, which is used to generate a highly uniform diffuse scattering light field in the incident angle range of 10°-80° and suppress specular reflection. The multi-view synchronous acquisition module is deployed around the automotive parts inspection station and includes at least three industrial cameras arranged at different elevation angles. It acquires multi-angle raw images with reference points by synchronously triggering the acquisition device. A scattering feature intelligent positioning module, signal-connected to the multi-view synchronous acquisition module, is used to convert physical scattering characteristics into mathematically solvable spatial coordinate data, specifically including: Scattering schlieren feature extraction unit: Based on the optical properties of the reference point microstructure, it separates the scattering region from the multi-view original image, calculates the radial brightness gradient distribution, and generates a feature vector; Sub-pixel registration unit: Through a pre-trained triple attention convolutional neural network, the feature vector is matched with the reference scattering schlieren template, and the center coordinates of the reference point with an accuracy better than 0.05 pixels are output. Multi-view geometry solving unit: Establishes matching point pairs based on the sub-pixel center coordinates, iteratively solves the camera extrinsic parameter matrix using the Zhang Zhengyou calibration method, and performs weighted RANSAC point cloud initial registration; The dynamic calibration and compensation module receives the initial registration results output by the multi-view geometry solution unit, corrects the scattering compensation coefficient of the reference point in the high curvature region in real time according to the surface curvature distribution, and outputs the corrected high-precision 3D reconstruction data.

2. The system for tracing and locating surface defects in automotive manufacturing parts according to claim 1, characterized in that: The isotropic scattering reference point module includes: Microcavity array unit: A conical microcavity with a sidewall tilt angle of θ = 55 ± 2° is formed by etching the surface of an alumina ceramic substrate, which is used to generate a uniform scattering field in the range of 10°-80° incident angle; Anti-reflective coating unit: A layer of silica nanoparticles coated on the surface of the microcavity, with titanium dioxide photocatalytic particles embedded on its surface to decompose the attached oil film and maintain scattering characteristics; Mechanical positioning unit: It has a built-in electromagnet and flexible universal joint, which makes the reference point adaptively fit the curved surface of the workpiece.

3. The system for tracing and locating surface defects in automotive manufacturing parts according to claim 1, characterized in that: The convolutional neural network is trained through the following steps: During the sample collection phase, curvature values ​​of 0-200 μm were covered on the surfaces of galvanized steel sheets, high-gloss polyurethane paint surfaces, and carbon fiber composite materials. -1 In the test area, millions of frames of reference point scattered light spot image samples were collected using multi-angle industrial cameras; In the feature engineering stage, a 16-direction radial brightness gradient histogram with the reference point center as the origin is extracted for each frame of scattered light spot image. The average gradient magnitude is calculated by dividing the angle interval into 0°-360° equally and combining them into a 128-dimensional schlieren feature vector. During the network optimization phase, a triple attention mechanism is constructed by using cascaded channel attention branches, spatial attention branches, and curvature attention branches to optimize feature matching.

4. The system for tracing and locating surface defects in automotive manufacturing parts according to claim 1, characterized in that: The sub-pixel registration unit includes: Curvature-responsive compensation mechanism: When the received real-time curvature K of the surface is greater than 50m -1 At that time, retrieve the refractive index η of the corresponding material; And activate the bilinear interpolation algorithm to calculate the edge refraction distortion compensation amount, the formula is: δ=0.02·|K|·(η-1); Corrected reference point center coordinates: (x', y') = (x + δ) x ,y+δ y ); In the formula: (x,y) are the coordinates of the center of the initially detected reference point; δ x and δ y Here are the distortion compensation amounts in the X and Y directions; (x', y') are the corrected real physical coordinates. δ represents the edge refraction distortion compensation amount, 0.02 is the curvature-refraction coupling coefficient, determined through ray tracing simulation and fitting with actual vehicle data; |K| is the absolute value of the surface curvature obtained by point cloud computing, with an effective range of 50-250m. -1 η is the relative refractive index of the workpiece, which is mapped through a pre-stored material database.

5. The surface defect tracing and location detection system for automotive manufacturing parts according to claim 4, characterized in that: The subpixel registration unit also includes an illumination fluctuation response mechanism: when the ambient illumination sensor detects a fluctuation amplitude >15%, the industrial camera is controlled to switch to HDR mode, and three layers of imaging are performed with three exposure times. The schlieren feature vectors of each exposure layer image are extracted and fused by weighting the inverse of the exposure time to output the subpixel center coordinates that are resistant to glare interference.

6. The surface defect tracing and location detection system for automotive manufacturing parts according to claim 1, characterized in that: The multi-view geometry solving unit is also used to perform thermal expansion compensation: Receive the workpiece surface temperature T collected by the infrared sensor and apply the deformation compensation matrix: [Δx,Δy]=[α·(T-T0)·L,β·(T-T0)·L] In the formula: α and β are the coefficients of thermal expansion of the workpiece material, which are pre-stored in the process database; T0 is the system calibration reference temperature; L is the Euclidean distance from the reference point to the nearest fixed edge of the workpiece, in mm; The initial registration coordinates are corrected to: (x', y') = (x + Δx, y + Δy) In the formula: (x,y) are the initial detected coordinates of the reference point center, which are subject to thermal deformation error; Δx and Δy are the compensation displacements caused by thermal expansion, calculated by the deformation matrix; (x',y') are the corrected real physical coordinates.

7. The system for tracing and locating surface defects in automotive manufacturing parts according to claim 1, characterized in that: In the dynamic calibration and compensation module, the formula for calculating the scattering compensation coefficient in real time based on the workpiece material type and local surface curvature is as follows: γ=1+λ×|K| In the formula: λ is the dynamic material coefficient, which is determined by mapping through a pre-stored database; |K| is the absolute value of the surface curvature, which is calculated by the difference of the point cloud normal vector; 1 is the baseline, i.e., the plane; The compensation is performed as follows: the initial registered point cloud coordinates (x0, y0, z0) are corrected to (x, y, z) = γ·(x0, y0, z0).

8. The system for tracing and locating surface defects in automotive manufacturing parts according to claim 7, characterized in that: It also includes a closed-loop error convergence module, which is connected to the output of the dynamic calibration compensation module. The closed-loop error convergence module includes: The benchmark comparison unit is used to perform ICP matching between the corrected 3D reconstructed point cloud and the laser tracker benchmark point cloud to calculate the Hausdorff distance ε. The dynamic optimization unit for the λ value updates the material coefficient λ of weight 7 in reverse iteration when ε > 0.03 mm. l new =λ old +sign(Not)×0.0005 In the formula: λ old λ is the material coefficient used in the current batch. new The material coefficient for the next batch of optimization; Δε = current error - previous error, sign is the sign function; The historical error database construction unit stores the optimal λ value under different material and curvature combinations, forming a process knowledge graph; The convergence verification unit locks the current λ value and outputs a calibration completion signal when ε≤0.02mm is reached for 3 consecutive batches.

Images