3D Reconstruction Method of High-Pressure Molded Parts for Lightweight Automobiles

Abstract

This invention relates to the field of intelligent inspection and 3D visual reconstruction technology for automotive parts, specifically a method for 3D image reconstruction of lightweight internal high-pressure molded parts for automobiles. The method includes: acquiring a sequence of structured light images, multispectral reflectance texture images, and a priori simulation model of the target object, and completing spatial coordinate registration; dynamically allocating edge extraction weights based on the local curvature gradient of the 2D image to generate a non-uniform density 3D point cloud; constructing a surface reconstruction energy function using surface reflectance variation data as a penalty term, fitting the point cloud to the 3D surface, and generating a target 3D mesh model; fusing 3D geometric features, 2D texture features, and the priori simulation model, and outputting the wall thickness reduction rate and residual stress distribution via a graph neural network; generating and overlaying a risk heat map, and outputting the evaluation decision results. This invention extends from geometric reconstruction to risk semantic reconstruction, improving the comprehensiveness and reliability of online inspection of internal high-pressure molded parts.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and three-dimensional vision reconstruction technology for automotive parts, specifically a method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automotive components. Background Technology

[0002] With the development of automotive lightweighting and internal high-pressure forming technology, internal high-pressure formed parts such as intake and exhaust pipes and high-precision flanges have become important objects in automotive parts manufacturing. In order to achieve quality control and service reliability assessment of such formed parts, it is particularly important to perform high-precision three-dimensional reconstruction of the workpiece surface morphology and combine it with defect risk analysis.

[0003] In the inspection of internally formed high-pressure parts, existing technologies mostly use structured light scanning, ordinary visual imaging, or single geometric measurement methods to inspect the shape of the workpiece. Although these methods can obtain the contour and size information of the workpiece to a certain extent, they often suffer from insufficient reconstruction accuracy, obvious artifact interference, and weak correlation between the inspection results and the actual physical state when dealing with situations such as high reflectivity of metal surfaces, drastic changes in local curvature, thinning of wall thickness after forming, and difficulty in directly characterizing residual stress. Especially in the context of online inspection on the production line, if only uniform point cloud reconstruction or single image information is relied upon, it is difficult to simultaneously take into account the reconstruction resolution of key areas, the ability to suppress surface optical interference, and the ability to accurately identify potential failure areas.

[0004] Therefore, how to collaboratively process the structured light image data, surface reflection texture data, and prior information of the molding process of internal high-pressure molded parts to obtain a three-dimensional reconstruction result that takes into account both geometric accuracy and physical state characterization ability, is crucial for the subsequent generation of visual evaluation results by combining risk information such as wall thickness reduction rate and residual stress, so as to build a corresponding online detection and quality judgment mechanism. This is essential for ensuring the manufacturing quality and reliability of automotive internal high-pressure molded parts. Summary of the Invention

[0005] The purpose of this invention is to provide a method for 3D image reconstruction of lightweight automotive internal high-pressure molded parts, addressing the following technical problems: Existing internal high-pressure molded part inspection technologies have significant shortcomings in 3D reconstruction accuracy when faced with high reflectivity of metal surfaces and drastic local curvature changes, as well as in characterizing and correlating actual physical states such as wall thickness reduction and residual stress based on surface morphology. There is an urgent need to propose a method for 3D image reconstruction of automotive lightweight internal high-pressure molded parts that can more accurately and collaboratively process multimodal image data and prior information about the molding process, dynamically predict internal physical states, and generate visualized evaluation results. The purpose of this invention can be achieved through the following technical solutions:

[0006] A method for 3D reconstruction of images of lightweight automotive internal high-pressure molded parts includes the following steps:

[0007] S1. Obtain the structured light image sequence and multispectral reflectance texture image of the target object, and obtain the spatial calibration parameters for multimodal data alignment. Based on the spatial calibration parameters, perform spatial coordinate registration on the structured light image sequence and multispectral reflectance texture image, and obtain the prior simulation model of the target object.

[0008] S2. Calculate the local curvature gradient of a two-dimensional image based on the structured light image sequence, dynamically allocate edge extraction weights according to the local curvature gradient of the two-dimensional image, and generate a three-dimensional point cloud with non-uniform density.

[0009] S3. Extract the surface reflectance variation data of the multispectral reflectance texture image, use the surface reflectance variation data as a penalty term to construct the surface reconstruction energy function, and perform three-dimensional surface fitting on the three-dimensional point cloud based on the surface reconstruction energy function to generate the target three-dimensional mesh model.

[0010] S4. Extract the three-dimensional geometric features of the target three-dimensional mesh model and the two-dimensional texture features of the multispectral reflectance texture image. Construct an initial graph structure with the vertices of the target three-dimensional mesh model as graph nodes and the mesh topology edges of the target three-dimensional mesh model as graph edges. Map the three-dimensional geometric features, two-dimensional texture features and prior simulation model to the graph nodes. Calculate the feature difference between adjacent vertices and map it to the graph edges. Input the pre-trained graph neural network for feature fusion and mapping, and output the physical state distribution data representing the wall thickness reduction rate and residual stress.

[0011] S5. Generate a two-dimensional risk heat map based on physical state distribution data, and overlay the risk heat map onto the surface of the target three-dimensional mesh model based on spatial coordinate mapping to generate a multi-dimensional risk map. Output the assessment and decision results of the target object based on the multi-dimensional risk map.

[0012] Optionally, S2 includes the following sub-steps:

[0013] S21. Extract the fringe phase map from the structured light image sequence and analyze it to extract the two-dimensional image curvature data of the target object;

[0014] S22. Calculate the spatial rate of change of curvature data of a two-dimensional image to generate a local curvature gradient of the two-dimensional image;

[0015] S23. Determine the relationship between the local curvature gradient of a two-dimensional image and a preset curvature gradient threshold;

[0016] S24. If the local curvature gradient of the two-dimensional image is greater than or equal to the preset curvature gradient threshold, then the first edge extraction weight is assigned, and a high-density point cloud region is generated based on the first edge extraction weight.

[0017] S25. If the local curvature gradient of the two-dimensional image is less than the preset curvature gradient threshold, then a second edge extraction weight is assigned, and a low-density point cloud region is generated based on the second edge extraction weight, wherein the first edge extraction weight is greater than the second edge extraction weight, so as to combine and generate a three-dimensional point cloud with non-uniform density.

[0018] Optionally, S3 includes the following sub-steps:

[0019] S31. Extract pixel intensity distribution information from multispectral reflectance texture images;

[0020] S32. Calculate the reflectance gradient between adjacent pixels based on pixel intensity distribution information to generate surface reflectance change data;

[0021] S33. Normalize the surface reflectivity variation data to generate depth correction coefficients;

[0022] S34. Embed the depth correction coefficient as a penalty term into the preset basic surface reconstruction function, wherein the basic surface reconstruction function includes a data term that constrains the three-dimensional mesh vertices to approximate the original point cloud and a smoothing term that constrains the continuity of the local surface, so as to construct the surface reconstruction energy function.

[0023] S35. Run the surface reconstruction energy function to perform 3D surface fitting and spatial coordinate iterative correction on the 3D point cloud to suppress surface gloss artifacts and output the target 3D mesh model.

[0024] Optionally, S4 includes the following sub-steps:

[0025] S41. Analyze the vertex coordinates and topological relationships of the target 3D mesh model, and compare them spatially with the standard computer-aided design model in the prior simulation model to extract the local curvature and deformation gradient of the 3D mesh, and combine the local curvature and deformation gradient of the 3D mesh into 3D geometric features.

[0026] S42. Perform frequency domain transformation on the local multispectral reflectance texture image region that matches the spatial dimensions of each vertex of the target 3D mesh model to obtain frequency domain information, analyze the frequency domain information to extract the surface texture anisotropic features and surface roughness that characterize the material deformation, and combine the surface texture anisotropic features and surface roughness into the two-dimensional texture features of the corresponding graph nodes.

[0027] S43. Analyze the prior simulation model to extract the initial water expansion pressure and initial wall thickness distribution data;

[0028] S44. Perform multi-dimensional feature stitching on the three-dimensional geometric features, two-dimensional texture features, and initial water expansion pressure and initial wall thickness distribution data to generate the fused feature vector of the corresponding graph node; at the same time, calculate the difference between the three-dimensional geometric features, two-dimensional texture features, initial water expansion pressure and initial wall thickness distribution data between adjacent vertices to generate the edge feature vector of the corresponding graph edge.

[0029] S45. Input the fused feature vector and edge feature vector into the pre-trained graph neural network for forward propagation to output physical state distribution data.

[0030] Optionally, the pre-trained graph neural network is generated through the following steps:

[0031] Acquire the three-dimensional geometric features, two-dimensional texture features, and prior simulation sample data of historical samples that are the same model and use the same molding process as the target object;

[0032] The actual wall thickness reduction rate and actual residual stress distribution obtained from historical samples through destructive testing are used as label data;

[0033] Construct the initial graph neural network;

[0034] The three-dimensional geometric sample features, two-dimensional texture sample features, and prior simulation sample data are input into the initial graph neural network for prediction to generate predicted state distribution data.

[0035] Calculate the loss value between the predicted state distribution data and the label data;

[0036] The network parameters of the initial graph neural network are updated using the backpropagation algorithm based on the loss value until the loss value converges, thereby generating a pre-trained graph neural network.

[0037] Optionally, S5 includes the following sub-steps:

[0038] S51. Obtain the preset forming limit curve and yield strength threshold of the target material, substitute the wall thickness reduction rate and residual stress values ​​in the physical state distribution data into the preset nonlinear material failure evaluation function for joint solution, so as to obtain the risk value characterizing the margin from the failure boundary, and match the preset color mapping dictionary according to the risk value to generate a two-dimensional risk heat map.

[0039] S52. Based on spatial coordinate mapping, the pixels of the risk heat map are mapped to the corresponding vertices of the target 3D mesh model to generate a 3D multi-dimensional risk map, wherein the multi-dimensional risk map contains the risk value of each vertex;

[0040] S53. Extract the maximum value of the risk value of each vertex in the multidimensional risk map as the highest risk value;

[0041] S54. Determine the relationship between the highest risk value and the preset failure threshold;

[0042] S55. If the highest risk value is greater than or equal to the failure threshold, the evaluation decision result is output as unqualified.

[0043] S56. If the highest risk value is less than the failure threshold, the evaluation decision result is output as qualified.

[0044] Optionally, the target object is a high-pressure molded part for automotive interiors, wherein the high-pressure molded part for automotive interiors includes intake and exhaust pipes or high-precision flanges.

[0045] Optionally, the prior simulation model includes a standard computer-aided design model and molding simulation data, wherein the molding simulation data includes initial water expansion pressure and initial wall thickness distribution data.

[0046] Optionally, the multispectral reflectance texture image is acquired using a high-resolution polarization camera, and the structured light image sequence is acquired using a laser stripe scanner.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This method achieves the fusion of geometric information, surface optical information and molding process prior information by collaboratively acquiring, spatially registering and uniformly processing the structured light image sequence, multispectral reflective texture image and prior simulation model of the target object. It can extend the traditional detection method that only focuses on the shape contour to a semantic-level reconstruction mechanism that simultaneously focuses on the wall thickness reduction rate and residual stress distribution, thereby enhancing the correlation between online detection results and actual physical state.

[0049] 2. This method dynamically allocates edge extraction weights based on the local curvature gradient of a 2D image, forming high-density point clouds in areas with drastic curvature changes such as flange transition fillets and the outer arc of bends, while maintaining low-density point clouds in straight areas. This improves the reconstruction resolution of key risk areas while keeping the overall data volume and computational resources under control, avoiding the problem of insufficient attention to important areas in uniform sampling methods. By extracting surface reflectance variation data from multispectral reflectance texture images and embedding it as a penalty term into the surface reconstruction energy function, the 3D point cloud is iteratively corrected. This effectively suppresses depth artifacts such as false bumps and false depressions caused by high metallic reflectivity, local brilliance, and uneven specular reflection, improving the surface fitting accuracy and physical reliability of the target 3D mesh model.

[0050] 3. This method extracts three-dimensional geometric features from the target three-dimensional mesh model and two-dimensional texture features from the multispectral reflectance texture image. It then combines a standard computer-aided design model with molding simulation data containing initial water expansion pressure and initial wall thickness distribution data to construct a graph structure input pre-trained graph neural network. This enables the establishment of a mapping relationship between surface morphology, surface texture and internal physical state, thereby achieving non-destructive prediction of wall thickness reduction rate and residual stress distribution, and improving the ability to identify potential failure areas.

[0051] 4. This method generates a two-dimensional risk heat map based on physical state distribution data and overlays it onto the surface of the target three-dimensional mesh model to form a multi-dimensional risk map. At the same time, it combines the failure critical threshold to output the evaluation decision results of qualified, unqualified or pending review. This enables quality inspectors to intuitively locate high-risk areas and supports automatic sorting, re-inspection guidance and online release judgment on the production line, thereby improving the accuracy and effectiveness of online quality control and service reliability assessment of high-pressure molded parts in automobiles. Attached Figure Description

[0052] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0053] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0055] like Figure 1 As shown, the method for 3D reconstruction of images of high-pressure molded parts for lightweight automobiles includes the following steps:

[0056] S1. Obtain the structured light image sequence and multispectral reflectance texture image of the target object, and obtain the spatial calibration parameters for multimodal data alignment. Based on the spatial calibration parameters, perform spatial coordinate registration on the structured light image sequence and multispectral reflectance texture image, and obtain the prior simulation model of the target object.

[0057] S2. Calculate the local curvature gradient of a two-dimensional image based on the structured light image sequence, dynamically allocate edge extraction weights according to the local curvature gradient of the two-dimensional image, and generate a three-dimensional point cloud with non-uniform density.

[0058] S3. Extract the surface reflectance variation data of the multispectral reflectance texture image, use the surface reflectance variation data as a penalty term to construct the surface reconstruction energy function, and perform three-dimensional surface fitting on the three-dimensional point cloud based on the surface reconstruction energy function to generate the target three-dimensional mesh model.

[0059] S4. Extract the three-dimensional geometric features of the target three-dimensional mesh model and the two-dimensional texture features of the multispectral reflectance texture image. Construct an initial graph structure with the vertices of the target three-dimensional mesh model as graph nodes and the mesh topology edges of the target three-dimensional mesh model as graph edges. Map the three-dimensional geometric features, two-dimensional texture features and prior simulation model to the graph nodes. Calculate the feature difference between adjacent vertices and map it to the graph edges. Input the pre-trained graph neural network for feature fusion and mapping, and output the physical state distribution data representing the wall thickness reduction rate and residual stress.

[0060] S5. Generate a two-dimensional risk heat map based on physical state distribution data, and overlay the risk heat map onto the surface of the target three-dimensional mesh model based on spatial coordinate mapping to generate a multi-dimensional risk map. Output the assessment and decision results of the target object based on the multi-dimensional risk map.

[0061] This embodiment provides a three-dimensional image reconstruction mechanism for lightweight internal high-pressure formed parts in automobiles. Specifically, the following description uses an internal high-pressure water-expanded stainless steel bend in an automobile exhaust system as the main scenario. After the bend is formed on the production line, it is subjected to structured light scanning, polarized multispectral reflectance texture imaging, three-dimensional reconstruction, wall thickness and residual stress prediction, and qualification judgment by an online inspection station. Subsequent implementation methods all revolve around different processing stages of the same bend in the same inspection process.

[0062] In this paper, the structured light image sequence is specifically formed by a laser stripe scanner. Therefore, the laser stripe image, stripe image, and structured light image mentioned in the descriptive description all refer to the same type of input data used for geometric reconstruction. Correspondingly, the multispectral reflectance texture image includes surface texture images obtained under different wavelengths and polarization states, which are used to characterize surface optical reflection and directional texture information. For example, the surface reflectance variation data in this paper is a general term for the constraints involved in reconstruction in step S3. In implementation, it can be calculated from the reflectance gradient, reflectance gradient, or equivalent pixel intensity change between adjacent pixels. The depth correction coefficient is the result after normalizing the surface reflectance variation data. The two correspond to each other, have different meanings, and are not used interchangeably.

[0063] Specifically, in step S1, the detection station first acquires a structured light image sequence and a multispectral reflection texture image of the target object; the structured light image sequence can be understood as several frames of stripe images taken from different perspectives of the workpiece, and the multispectral reflection texture image records the reflection differences of the workpiece surface in different bands and polarization states; the purpose of spatial coordinate registration is to unify the image of the acquisition depth with the image of the acquisition surface optical features into the same workpiece coordinate system; schematically, it is assumed that the laser stripe scanner acquires three spatial points P1, P2, and P3 on the workpiece surface, with coordinates (10,5,2), (12,5,3), and (14,6,3) respectively, and the high-resolution polarization camera records three corresponding pixels q1, q2, and q3 in the pixel plane, with pixel coordinates (101,205), (115,206), and (129,212) respectively;

[0064] After calibration matrix and extrinsic parameter transformation, a one-to-one mapping relationship can be established between q1, q2, q3 and P1, P2, P3, so that any subsequent three-dimensional point can be indexed to its corresponding multispectral reflectance information; at the same time, the system loads the prior simulation model of the corresponding model of the bend, which can include the standard computer-aided design model and the simulation results of the initial wall thickness distribution, initial pressure distribution, etc. of the model under the given water expansion pressure process;

[0065] In step S2, the system does not use full-area equal-precision sampling. Instead, it first estimates the local curvature gradient of the two-dimensional image from the structured light image sequence, and then dynamically assigns edge extraction weights accordingly. For example, suppose the local image after the pipe is unfolded is divided into three neighborhoods: A, B, and C. A corresponds to the flange transition fillet, B corresponds to the outer arc of the pipe, and C corresponds to a relatively straight area. After coarse surface restoration, the two-dimensional local curvature values ​​of A, B, and C can be recorded as 0.82, 0.56, and 0.08, respectively. After changes in adjacent positions, the curvature gradients are 0.31, 0.22, and 0.03, respectively. The system will assign higher edge extraction weights to A and B, and lower weights to C.

[0066] Therefore, denser edge and contour points are retained in regions A and B, while the point cloud density is appropriately reduced in region C. The resulting three-dimensional point cloud is not a traditional uniform point cloud, but a non-uniform density point cloud adapted to the failure-sensitive region.

[0067] In step S3, the system further extracts the surface reflectance variation data from the multispectral reflectance texture image and adds it as a penalty term to the surface reconstruction energy function to reduce depth artifacts caused by metallic highlights, local brightening, and uneven specular reflection. For ease of understanding, it is assumed that there are three regions to be fitted in the point cloud: R1, R2, and R3, where R2 is located in the brighter polished area on the outer arc of the curved pipe. If only the smoothness of the geometric neighborhood is used for fitting, R2 may be fitted as a false local bulge due to reflection.

[0068] At this point, the system statistically analyzes the changes in multispectral pixel intensity in region R2 and finds that its reflectance gradient is significantly higher than that of R1 and R3. Therefore, a higher penalty coefficient is applied to the depth correction term of this region, forcing the fitted surface to satisfy geometric continuity while avoiding mistaking simple gloss abrupt changes as real morphological abrupt changes, and finally outputting the target 3D mesh model. Furthermore, in the subsequent implementations of this paper, if reflectance gradient and reflectance gradient are mentioned, they are specific calculation expressions of surface reflectance change data, without changing their unified meaning as input quantities in step S3.

[0069] In step S4, the system extracts three-dimensional geometric features from the target three-dimensional mesh model and two-dimensional texture features from the multispectral reflectance texture image. These features are then input into a pre-trained graph neural network along with the prior simulation model to obtain physical state distribution data characterizing the wall thickness reduction rate and residual stress. The graph structure here can be illustrated as follows: Assume a local mesh consists of four vertices V1, V2, V3, and V4, with edge relationships of V1-V2, V2-V3, V3-V4, and V1-V4. The system uses vertices as graph nodes and edges as graph edges. It combines the surface two-dimensional curvature data and deformation gradient of each vertex into three-dimensional geometric features and maps them to the graph nodes. It also combines the grain stretching direction and surface roughness of the corresponding projection region of the vertex into two-dimensional texture features and maps them to the same graph node. Finally, it maps the initial water expansion pressure and initial wall thickness distribution in the prior simulation model to the initial state of the nodes.

[0070] Meanwhile, to maintain consistency with the representation mapped to graph nodes and edges, the system can also write the curvature difference, deformation gradient difference, texture principal direction angle difference, and prior wall thickness difference between adjacent vertices into the corresponding graph edges to represent the strength of local state transitions. After multi-layer propagation through the graph neural network, the network outputs a set of state values ​​for each vertex. For example, V1 has a wall thickness reduction rate of 8% and a residual stress of 120MPa; V2 has 13% and 155MPa; V3 has 21% and 240MPa; and V4 has 11% and 130MPa. This forms the physical state distribution data for the entire workpiece surface.

[0071] In step S5, the system generates a two-dimensional risk heat map based on the physical state distribution data, and then overlays it onto the surface of the target three-dimensional mesh model through spatial coordinate mapping to form a multi-dimensional risk map and output the evaluation decision results; schematically, the system can weight the wall thickness reduction rate and residual stress, for example, the risk value can be expressed as 0.6×normalized wall thickness reduction rate + 0.4×normalized residual stress;

[0072] Using the example of the four vertices mentioned above, if the comprehensive risk value of V3 is the highest and reaches the preset threshold, it will be displayed as a red area on the two-dimensional heat map, and a high-risk label will be formed at the outer arc transition position of the three-dimensional pipe model; the evaluation decision result can be output as qualified or unqualified, and a suggested re-inspection area will be given.

[0073] Furthermore, in step S1, if the structured light image at a certain viewpoint is overexposed or occluded, the viewpoint can be marked as a low-confidence frame and will not participate in the main reconstruction, but will only be used as an auxiliary reference; if the reprojection error after the registration of the texture image and the structured light image exceeds the set threshold, the calibration parameters will be called again or local reregistration will be performed; in step S2, if the local curvature gradient calculation result is generally low, it may indicate that there is contamination on the surface of the current workpiece, resulting in unclear edges. At this time, the system can trigger a second exposure acquisition or use curvature estimation after neighborhood smoothing to avoid excessive sparseness of the point cloud;

[0074] In step S3, if the reflectance change in a certain area is abnormally large but inconsistent with the multi-view geometry, it is preferentially identified as an optical artifact rather than deformation; if the reflectance change and geometric change are consistent in multiple frames of images, it is retained as a true feature; in step S4, if the local wall thickness reduction rate output by the graph neural network is negative or exceeds the process limit, the result is included in the abnormal node set, and the outputs of adjacent nodes are combined for smoothing constraints or back to the prior simulation value; in step S5, if the highest risk value is close to the threshold boundary, for example, only differing by one tolerance interval, it is output as pending review rather than directly judged as qualified or unqualified, in order to avoid misjudgment of boundary samples;

[0075] For example, in the online inspection of exhaust pipes, the system completes multi-modal acquisition of the outer arc, inner arc, and flange connection of the pipe; forms a high-density point cloud at the flange transition fillet and the outer arc of the pipe, while maintaining a lower density in the straight pipe section; uses the change in texture reflectivity to correct the local geometric reconstruction deviation caused by specular reflection; then uses a graph neural network to transform the implicit correspondence between morphology, texture, and forming simulation into wall thickness reduction rate and residual stress distribution; outputs a risk heat map on the three-dimensional pipe surface, and determines whether the part meets the conditions for subsequent vehicle installation and use.

[0076] The purpose of this step is to expand simple geometric reconstruction into semantic-level reconstruction oriented towards durability and sealing reliability, so that the system output includes not only the surface geometric features of the workpiece, but also the internal physical state and failure risk distribution of the workpiece, thereby improving the comprehensiveness and reliability of online inspection of internal high-pressure forming parts.

[0077] In this embodiment, S2 includes the following sub-steps:

[0078] S21. Extract the fringe phase map from the structured light image sequence and analyze it to extract the two-dimensional image curvature data of the target object;

[0079] S22. Calculate the spatial rate of change of curvature data of a two-dimensional image to generate a local curvature gradient of the two-dimensional image;

[0080] S23. Determine the relationship between the local curvature gradient of a two-dimensional image and a preset curvature gradient threshold;

[0081] S24. If the local curvature gradient of the two-dimensional image is greater than or equal to the preset curvature gradient threshold, then the first edge extraction weight is assigned, and a high-density point cloud region is generated based on the first edge extraction weight.

[0082] S25. If the local curvature gradient of the two-dimensional image is less than the preset curvature gradient threshold, then a second edge extraction weight is assigned, and a low-density point cloud region is generated based on the second edge extraction weight, wherein the first edge extraction weight is greater than the second edge extraction weight, so as to combine and generate a three-dimensional point cloud with non-uniform density.

[0083] This embodiment provides a curvature gradient driving mechanism for generating non-uniform density three-dimensional point clouds; specifically, this embodiment follows the processing procedure of the same exhaust pipe in step S2 mentioned above, and is used to further illustrate how to determine high-risk areas from structured light image sequences and adopt higher sampling and edge extraction intensities.

[0084] Specifically, if we only say that the weights are dynamically allocated based on the curvature gradient at the level of the embodiment, in some cases where the transition rounded corners are small and the stripe contrast is unstable, the problem of inconsistent weight allocation may still occur. That is, adjacent positions in the same high-risk area may be judged as high density in one position and low density in another position, resulting in local breakage of the point cloud. Therefore, this embodiment refines step S2 into a set of executable sub-steps.

[0085] In step S21, the system analyzes the structured light image sequence to extract the surface two-dimensional curvature data of the target object; the two-dimensional curvature data can be derived from the fringe phase map, the degree of contour line curvature, or the curvature approximation value converted from the coarse depth map in the image plane.

[0086] Specifically, when using a fringe phase map to extract two-dimensional curvature, the system performs phase shift calculations and phase unrolling on the fringe image to obtain a continuous absolute phase distribution. The magnitude of the two-dimensional principal curvature is approximated using second-order partial derivatives:

[0087]

[0088] in, and These represent the pixel coordinates of the two-dimensional fringe phase map in the horizontal and vertical directions, respectively, serving as a quantitative representation of the surface two-dimensional curvature data.

[0089] To illustrate, suppose a contour line is cut from near the outer arc of the bend, and the curvature values ​​of five consecutive sampling points on it are 0.10, 0.12, 0.40, 0.75 and 0.78 respectively. It can be seen that the contour line transitions from a gentle curve to a sharp bend.

[0090] In step S22, the system further calculates the spatial rate of change of curvature data to generate a local curvature gradient of a two-dimensional image. The difference between adjacent sampling points can be used for simplification. Taking the above five points as an example, the adjacent differences are 0.02, 0.28, 0.35, and 0.03. Therefore, the curvature gradient near the third and fourth positions is significantly higher than that of other positions. This usually corresponds to the dangerous boundary on the workpiece where the flat section transitions to the outer bulge section.

[0091] In step S23, the system determines the relationship between the local curvature gradient of the two-dimensional image and the preset curvature gradient threshold. Assuming the preset threshold is set to 0.20, then 0.28 and 0.35 belong to the region above the threshold, and 0.02 and 0.03 belong to the region below the threshold.

[0092] In step S24, if the local curvature gradient is greater than or equal to the threshold, a first edge extraction weight is assigned. Assuming the first weight is set to 0.9, the system will make the point retention rate of high gradient regions higher when performing sub-pixel edge localization, stripe center extraction or contour preservation. For example, on contour segments of equal length, high gradient regions retain 9 points, while low gradient regions retain only 4 points, thus forming a high-density point cloud region.

[0093] In step S25, if the local curvature gradient is less than the threshold, a second edge extraction weight is assigned; assuming the second weight is 0.4, the system reduces the edge sampling frequency and fitting priority of the region, retaining only the key points sufficient to maintain the overall contour; finally, after the high-density region and the low-density region are stitched together, a non-uniform density three-dimensional point cloud adapted to the complexity of the structure is formed.

[0094] Furthermore, if the curvature gradient is exactly equal to the threshold, it can be directly classified into a high-density region to avoid losing points at high-risk boundary locations; if the overall curvature gradient fluctuates very little, causing most regions to be judged as low-density, the system can activate the minimum point cloud density constraint, such as retaining at least the basic number of points per unit area to prevent excessive sparsity; if the curvature gradient of an isolated point is abnormally high due to noise, it is necessary to combine the neighborhood consistency judgment, and only when the point and its neighboring points have at least two consecutive positions that meet the threshold condition will the first weight be assigned to avoid mistaking noise points as feature regions;

[0095] For example, in the aforementioned online detection of the exhaust bend, the system discovers from the structured light image that the curvature gradient at the connection point between the flange root and the outer arc of the bend is continuously greater than a threshold. Therefore, a first edge extraction weight is assigned to this location to generate a high-density point cloud. For the straight section in the middle of the bend, since the curvature change is gradual, a second edge extraction weight is assigned, resulting in only a low-density point cloud. Thus, while ensuring data processing efficiency, the system concentrates more sampling capabilities on the parts most likely to experience wall thickness reduction and stress concentration.

[0096] The purpose of this step is to concentrate high-information sampling resources on areas with drastic curvature changes by using clear threshold judgment and dual weight allocation, thereby improving the reconstruction resolution of key parts while controlling the overall data volume.

[0097] In this embodiment, S3 includes the following sub-steps:

[0098] S31. Extract pixel intensity distribution information from multispectral reflectance texture images;

[0099] S32. Calculate the reflectance gradient between adjacent pixels based on pixel intensity distribution information to generate surface reflectance change data;

[0100] S33. Normalize the surface reflectivity variation data to generate depth correction coefficients;

[0101] S34. Embed the depth correction coefficient as a penalty term into the preset basic surface reconstruction function, wherein the basic surface reconstruction function includes a data term that constrains the three-dimensional mesh vertices to approximate the original point cloud and a smoothing term that constrains the continuity of the local surface, so as to construct the surface reconstruction energy function.

[0102] S35. Run the surface reconstruction energy function to perform 3D surface fitting and spatial coordinate iterative correction on the 3D point cloud to suppress surface gloss artifacts and output the target 3D mesh model.

[0103] This embodiment provides a texture-depth joint optimization mechanism for eliminating gloss artifacts on metal surfaces. Specifically, this embodiment follows the surface fitting stage of the same exhaust pipe after the non-uniform point cloud is generated, and focuses on how to use multispectral reflectance texture images to perform depth correction on the three-dimensional point cloud.

[0104] Specifically, while the non-uniform density point cloud generated by the embodiment can improve the geometric sampling accuracy of high curvature areas, it still has a prominent drawback for objects with obvious specular reflection characteristics, such as stainless steel workpieces: structured light or laser stripes may shift, saturate, or blur in bright areas, causing false depressions or bulges in local point clouds in three-dimensional space; these non-physical artifacts may seem geometrically reasonable, but their source is actually optical reflection rather than real deformation; therefore, this embodiment introduces a depth correction mechanism based on multispectral reflection texture.

[0105] In step S31, the system extracts pixel intensity distribution information from the multispectral reflectance texture image; a simplified explanation is as follows: assuming a local area is a 3×3 pixel block, its pixel intensity in a certain band is as follows: 80, 82, 85; 78, 140, 88; 76, 79, 81; among which, the central pixel 140 is much higher than the surrounding pixels, indicating that there is obvious reflectance enhancement at this location;

[0106] In step S32, the system calculates the reflectivity gradient between adjacent pixels based on the pixel intensity distribution information to form surface reflectivity change data; for example, taking the center pixel as a reference, the intensity differences between it and its four neighboring pixels can be 58, 52, 61, and 61, respectively, thus it can be considered that the center position is in a high reflectivity gradient region.

[0107] In step S33, the system normalizes the reflectivity change data to generate a depth correction coefficient. Schematic, if the minimum reflectivity gradient in all sampling areas of the current workpiece is 0 and the maximum is 80, then the reflectivity gradient 58 corresponding to the center pixel can be normalized to approximately 0.73. This 0.73 can be used as the subsequent depth correction coefficient, indicating that the point should be subject to strong constraints during depth fitting.

[0108] Furthermore, to avoid a few extreme bright spots from amplifying the normalization scale of the entire image, the system can prioritize normalization using the effective gradient statistics range within the current region or frame, and truncate gradient values ​​exceeding the upper boundary to ensure the depth correction coefficient stably falls within the 0 to 1 range. Thus, the business meaning of the depth correction coefficient can be clearly understood as follows: the closer the value is to 1, the more likely the location is to be affected by optical bright spots, and the lower the confidence in the original depth in subsequent iterations; the closer the value is to 0, the more the geometric measurement value is maintained.

[0109] In step S34, the system embeds the depth correction coefficient as a penalty term into a preset base surface reconstruction function to construct a surface reconstruction energy function; specifically, the surface reconstruction energy function... Define as a data item Smoothing terms With penalty items The weighted sum, i.e.

[0110]

[0111] Among them, data items 3D mesh vertices used for constraint fitting Approximating the original structured light point cloud coordinates Smoothing term Used to constrain the local surface continuity between adjacent vertices j; penalty term ,in, This refers to the depth correction coefficient generated in step S33. It is the unit vector in the direction of the line of sight; and These are the preset weight coefficients for the smoothing term and the penalty term, respectively; An independent index representing a 3D point cloud; Represents the vertices in the grid The set of adjacent topological nodes;

[0112] This formula shows that when the reflectivity at a certain location changes drastically, leading to... When the value approaches 1, the weight of the penalty term increases significantly, forcing the optimization process to reduce its reliance on the original depth artifact data at that point and instead rely mainly on the smoothing term to make geometric inferences from the neighborhood coordinates; that is, the larger the depth correction coefficient, the lower the system's confidence in the original depth value of that point.

[0113] Furthermore, the penalty term does not determine the final coordinates of a point alone, but rather participates in a trade-off with neighborhood continuity, multi-view consistency, and the local normal of the original point cloud. When the reflectivity of a point changes significantly but its geometric position is not consistent with that recovered from adjacent views, the penalty term prioritizes suppressing the abnormal offset of that point. When a highly reflective area repeatedly corresponds to the same stable morphological change under multiple views, the system only reduces its correction magnitude, rather than mechanically smoothing out the morphology completely.

[0114] In step S35, the system runs the aforementioned surface reconstruction energy function to iteratively correct the spatial coordinates of the 3D point cloud to eliminate surface gloss artifacts. A simplified deduction can be made: assuming the original depth of a point is 3.20mm, its neighborhood average depth is 3.00mm, and its corresponding depth correction coefficient is 0.73, then after the first iteration, the point can be corrected to 3.08mm; if the neighborhood consistency is further enhanced in the second iteration, it can continue to be corrected to 3.03mm; finally, the point no longer exhibits obvious local geometric reconstruction deviation, but returns to a position more consistent with the real surface; after all points are iterated, the target 3D mesh model is output; furthermore, during the iteration process, a single correction upper limit and stopping condition can be set, such as stopping the update when the correction amount in two consecutive rounds is lower than the preset tolerance, or outputting the current result when the preset maximum number of iterations is reached, to avoid the highlighted area being overly flattened or falling into invalid iterations;

[0115] Furthermore, if the overall pixel intensity of a certain area is low, causing the reflectivity gradient to be close to zero, then its depth correction coefficient is also close to zero. In this case, the system mainly relies on the geometric fitting term rather than the texture penalty term. If a certain area has abnormal intensity changes due to stains, oil films, or scratches, but the results from multiple viewing angles are inconsistent, the system can mark this area as a texture abnormality area and reduce its weight in depth correction to avoid mistaking surface contamination for gloss artifacts. If the maximum and minimum values ​​are the same during normalization, it indicates that the reflectivity of this batch of images changes very little. In this case, the depth correction coefficient can be directly set to the base value to avoid the denominator being zero or correction distortion.

[0116] For example, in the bright outer arc region of the same exhaust pipe, the local point cloud initially reconstructed by the structured light system shifts outward, forming a suspected bulge. After reading the multispectral reflectance texture image, the system found that there were obvious central bright and edge low bright features across the bands at this location, and the calculated reflectance gradient was significantly higher than that of the surrounding area. Therefore, the intensity of the penalty term in the surface reconstruction energy function was increased. After several iterations, the three-dimensional coordinates at this location fell back to a position that was consistent with the surrounding arc surface, and finally the mesh model that was more consistent with the surface of the real workpiece was output.

[0117] The purpose of this step is to convert the reflective information of the metal surface into a constraint quantity that can participate in three-dimensional fitting, so as to eliminate the shape artifacts caused by optical properties and improve the physical reliability of the reconstructed mesh.

[0118] In this embodiment, S4 includes the following sub-steps:

[0119] S41. Analyze the vertex coordinates and topological relationships of the target 3D mesh model, and compare them spatially with the standard computer-aided design model in the prior simulation model to extract the local curvature and deformation gradient of the 3D mesh, and combine the local curvature and deformation gradient of the 3D mesh into 3D geometric features.

[0120] S42. Perform frequency domain transformation on the local multispectral reflectance texture image region that matches the spatial dimensions of each vertex of the target 3D mesh model to obtain frequency domain information, analyze the frequency domain information to extract the surface texture anisotropic features and surface roughness that characterize the material deformation, and combine the surface texture anisotropic features and surface roughness into the two-dimensional texture features of the corresponding graph nodes.

[0121] S43. Analyze the prior simulation model to extract the initial water expansion pressure and initial wall thickness distribution data;

[0122] S44. Perform multi-dimensional feature stitching on the three-dimensional geometric features, two-dimensional texture features, and initial water expansion pressure and initial wall thickness distribution data to generate the fused feature vector of the corresponding graph node; at the same time, calculate the difference between the three-dimensional geometric features, two-dimensional texture features, initial water expansion pressure and initial wall thickness distribution data between adjacent vertices to generate the edge feature vector of the corresponding graph edge.

[0123] S45. Input the fused feature vector and edge feature vector into the pre-trained graph neural network for forward propagation to output physical state distribution data.

[0124] This embodiment provides a multi-dimensional feature fusion mechanism for predicting wall thickness reduction rate and residual stress; specifically, this implementation follows the physical state inversion stage of the same exhaust bend after obtaining the target three-dimensional mesh model.

[0125] Specifically, the target 3D mesh model obtained solely based on the embodiment, although its geometric shape is relatively accurate, still has a drawback: the geometric mesh can only show the spatial distribution of the surface, but cannot directly indicate what potential failure risks exist in that area; for example, two locations may have similar curvatures, but one location has higher internal residual stress and thinner wall thickness due to more severe material stretching during molding; without introducing texture and process priors, the system cannot distinguish between these two types of locations.

[0126] Therefore, this embodiment further integrates three-dimensional geometric features, two-dimensional texture features, and process features from the prior simulation model, and outputs the physical state distribution through a graph neural network. To avoid ambiguity in terminology, the three-dimensional geometric features here specifically refer to the local curvature and deformation gradient extracted from the target three-dimensional mesh model; the two-dimensional texture features specifically refer to the grain stretching direction and surface roughness extracted from the multispectral reflectance texture image; and the fused feature vector refers to the node feature encoding written to the graph node, which is not confused with the edge features written to the graph edge.

[0127] In step S41, the system analyzes the node coordinates and topological relationships of the target 3D mesh model, and uses the CAD model in the prior simulation model as the initial state reference to extract the local curvature and deformation gradient of the 3D mesh. A simplified illustration can be made: assuming the local mesh contains vertices V1, V2, and V3, with V2 located between V1 and V3, and the three forming an outer arc; if the normal change of V2 relative to V1 and V3 is more obvious, then V2's local curvature can be considered larger; if the offset of V2 from the corresponding position of the CAD standard model is more significant, then its deformation gradient is also larger; the system concatenates the local curvature and deformation gradient to form the 3D geometric features of the vertex, for example, the geometric features of V2 can be denoted as 0.72, 0.41;

[0128] In step S42, the system performs frequency domain transformation on the local multispectral reflectance texture image region that matches the vertex space of the 3D mesh model. After obtaining the frequency domain information, it extracts the surface texture anisotropy features and surface roughness. To simplify, if a local texture has obvious energy concentration along the 45-degree direction in the frequency domain, it indicates that the surface texture has strong directionality, which can correspond to the main flow extension of the material in this direction during the forming process. This direction is extracted as the surface texture anisotropy feature. If the total amount of high-frequency components is large, the surface roughness can be considered high. Assuming that the texture feature corresponding to the V2 projection region is 45°, 0.63, it indicates that its main deformation texture direction is obvious and the roughness is moderate to high, thus providing physical correlation clues for subsequent prediction of residual stress and thinning trend.

[0129] In step S43, the system analyzes the prior simulation model and extracts the initial water expansion pressure and initial wall thickness distribution data. For example, for the same bend, the simulation model can give the local initial water expansion pressure of 82MPa and the initial wall thickness of 1.50mm at the corresponding process stage V2. This data is not the final measured value, but the prior input of process knowledge.

[0130] In step S44, the system performs multi-dimensional feature splicing on the three-dimensional geometric features, two-dimensional texture features, and initial water expansion pressure and initial wall thickness distribution data to form a fused feature vector. Using the schematic data of V2, the fused feature vector can be represented as 0.72, 0.41, 45°, 0.63, 82, 1.50. In actual implementation, the angle can be further converted into sine and cosine forms to facilitate network learning, but its essence is still to uniformly encode features from different sources onto nodes or edges. Furthermore, to maintain consistency with the graph node and graph edge mapping in the embodiment, in addition to writing the above-mentioned fused features on the nodes, the system can also write edge features on the mesh topology edges. The edge features can be composed of the curvature difference, deformation gradient difference, texture principal direction angle difference, and prior wall thickness difference between adjacent vertices.

[0131] Schematic, if edge E12 connects V1 and V2, and the curvature difference between the two points is 0.18, the difference in the fringe angle is 12 degrees, and the initial wall thickness difference is 0.05 mm, then this edge can characterize the degree of abrupt transition of the local state from V1 to V2. The edge features here do not change the description of generating the fused feature vector in step S44. In specific implementation, it can be understood as follows: step S44 generates the fused feature vector on the node side, while the edge side features are derived from the same batch of basic features and synchronously written into the graph edge. The two together constitute the input of the graph neural network. In this way, when the graph neural network propagates, it not only knows what state a certain point itself is, but also knows how it changes between adjacent points.

[0132] In step S45, the system inputs the fused feature vector into a pre-trained graph neural network for forward propagation to output physical state distribution data; to achieve effective fusion of 3D heterogeneous features, the pre-trained graph neural network adopts a message passing mechanism based on edge feature awareness; for the ... Layer mesh vertices The iterative update formula for its node state is:

[0133]

[0134] in, For nodes In the The hidden state vector of the layer, the initial state of the input layer of the network That is, the fused feature vector. For nodes The set of topological neighbor nodes, and Here is the learnable weight matrix of the network. It is a non-linear activation function; in order to introduce process gradient features, aggregate weights are used. It incorporates local transition information from graph edges, and its calculation formula is as follows:

[0135]

[0136] in, To write edge features into the graph edges, It is a multilayer perceptron; Adjacent nodes In the Hidden state vector of layer; symbol This represents the vector feature concatenation operation; the role of graph neural networks is to propagate information about adjacent nodes using grid topological relationships.

[0137] For example, if V2 itself shows high risk, and its neighboring nodes V1 and V3 also have similar stretching directions and high curvature, the network will strengthen the judgment that V2 is in a weak zone; conversely, if V2 has high local curvature, but the surrounding nodes have smooth texture and low prior process pressure, the network may reduce its risk assessment.

[0138] Schematic, after forward propagation, the wall thickness reduction rate of V2 can be output as 18%, and the residual stress is 210MPa. Furthermore, the role of edge features in propagation is reflected in the adjustment of information transmission intensity: when the edge features between adjacent nodes indicate a continuous state transition, the network tends to enhance neighborhood consistency; when the edge features indicate the existence of abrupt boundary changes, the network retains local peaks to avoid over-smoothing the real high-risk concentration area.

[0139] Furthermore, if some mesh vertices fail to map to the effective texture region due to occlusion, their texture features can be set to default values, and the neighborhood propagation weight can be increased to compensate with the help of surrounding nodes; if a certain workpiece model lacks a corresponding complete prior simulation model, the system can degenerate to using only the standard CAD model and historical average process parameters to form a simplified prior input; if the local frequency domain transformation result does not have a significant principal direction peak, it indicates that the grain stretching direction in this region is not significant. In this case, its stretching direction can be marked as having no principal direction, and the roughness features should be retained; if the graph neural network output shows unreasonable jumps between adjacent nodes, post-processing smoothing can be performed in combination with the mesh topology, but the real high-risk peaks exceeding the threshold will not be changed;

[0140] For example, at the flange root of the same exhaust bend, the system detected high local curvature and deformation gradient in the 3D mesh; extracted from the multispectral reflectance texture image that the grain texture is concentrated along the outward pulling direction and the roughness increases; and found from the process simulation that this part experiences high internal pressure and large material flow during water expansion molding; the system stitches these features together and inputs them into a graph neural network, and finally outputs the high wall thickness reduction rate and residual stress value of the region, thus providing a more reliable physical basis for subsequent risk assessment.

[0141] The purpose of this step is to integrate the three heterogeneous types of information—geometry, texture, and process priors—into graph structure reasoning and establish a mapping relationship between surface representation and internal physical state.

[0142] In this embodiment, the pre-trained graph neural network is generated through the following steps:

[0143] Acquire the three-dimensional geometric features, two-dimensional texture features, and prior simulation sample data of historical samples that are the same model and use the same molding process as the target object;

[0144] The actual wall thickness reduction rate and actual residual stress distribution obtained from historical samples through destructive testing are used as label data;

[0145] Construct the initial graph neural network;

[0146] The three-dimensional geometric sample features, two-dimensional texture sample features, and prior simulation sample data are input into the initial graph neural network for prediction to generate predicted state distribution data.

[0147] Calculate the loss value between the predicted state distribution data and the label data;

[0148] The network parameters of the initial graph neural network are updated using the backpropagation algorithm based on the loss value until the loss value converges, thereby generating a pre-trained graph neural network.

[0149] This embodiment provides a training and generation mechanism for graph neural networks. Specifically, this implementation still revolves around the aforementioned exhaust bend detection line, but in terms of time dimension, it traces back to the offline modeling stage of this type of bend in historical batches. That is to say, the pre-trained graph neural network called by the current online detection is not directly generated without prior parameters, but is obtained by training together through the geometry, texture, simulation and destructive detection results of historical samples.

[0150] Specifically, although the embodiment provides an idea of ​​using fused features as input to a graph neural network for prediction, the mapping relationship may be too abstract if the method of obtaining the network is not explained. Especially in the case of internal high-pressure formed parts, the wall thickness reduction rate and residual stress cannot be directly read from a single surface imaging and must rely on historical samples to establish a supervised learning relationship. Therefore, this embodiment further discloses the network training process.

[0151] The system acquires historical sample features of the same type as the target object; assuming there are historical samples H1, H2, and H3, all of which are exhaust bends of the same model; each historical sample contains three-dimensional geometric sample features, two-dimensional texture sample features, and prior simulation sample data; for example, the input features of a local node of H1 can be curvature 0.70, deformation gradient 0.36, texture 30°, roughness 0.52, initial pressure 80, and initial wall thickness 1.50; H2 and H3 correspond to other process fluctuation samples;

[0152] The actual wall thickness reduction rate and actual residual stress distribution obtained from historical samples through destructive testing are used as label data; destructive testing can be obtained through methods such as slice thickness measurement, metallographic analysis, and residual stress testing; schematically, the actual label of the above node H1 may be a thinning rate of 17% and a residual stress of 205 MPa.

[0153] Construct an initial graph neural network; the network contains at least an input layer, several graph convolutional or graph messaging layers, and an output layer; the input layer receives the fused features of each grid node, the graph propagation layer transmits information between neighboring nodes according to the topological edges, and the output layer generates two types of values ​​for each node, namely the predicted value of the wall thickness reduction rate and the predicted value of the residual stress.

[0154] The training samples are input into the initial graph neural network for prediction, generating predicted state distribution data. For microscopic explanation, suppose that the first predicted output of a certain node H1 is a thinning rate of 12% and a residual stress of 170MPa, while the true label is 17% and 205MPa. This indicates that the network is currently underestimating the risk.

[0155] Furthermore, the system calculates the loss value between the predicted state distribution data and the label data; the loss value can be understood as the total deviation between the prediction and the reality; using the above example, the thinning rate deviation is 5 percentage points, and the residual stress deviation is 35 MPa; the system will accumulate the deviations of all nodes in the entire sample image to obtain the loss value of this round of training.

[0156] Based on this loss value, the network parameters are updated using the backpropagation algorithm until the loss value converges. To simplify, the loss value is 1.00 in the first training round, 0.73 in the second round, and 0.54 in the third round. If training continues until the tenth round, the loss value drops to 0.18, and the eleventh and twelfth rounds are 0.179 and 0.178 respectively, which are close to stabilization. Therefore, the loss value can be considered to have converged. At this point, the pre-trained graph neural network is obtained and deployed to the online detection station for use in the implementation example.

[0157] Furthermore, if the number of historical samples is insufficient, the system can adopt cross-validation and data augmentation strategies, such as treating different local areas of the same workpiece as sub-image samples to alleviate the scarcity of training samples; if some historical samples lack destructive detection labels, they can be used for unlabeled pre-training or directly removed to avoid contaminating the supervision signal; if the loss value first decreases and then increases significantly during training, it indicates that overfitting may occur, and an early stopping mechanism for the validation set can be introduced; if the dimensions of different labels are significantly different, such as the thinning rate being a percentage while the stress is in megapascals, normalization should be performed before calculating the loss to avoid the stress term unilaterally dominating the optimization direction;

[0158] For example, during the development phase of the aforementioned exhaust bend of the same model, the company accumulated scanned images, simulation process data, and actual measurement results of multiple batches of workpieces. The system uses these historical samples to build a training set, aligning the mesh nodes of each workpiece with its corresponding actual wall thickness reduction rate and actual residual stress. After multiple rounds of training, the network fits the nonlinear mapping law between high curvature, specific texture, and high initial pressure with high thinning rate and high residual stress. When a new online workpiece enters the inspection station, the trained network can be called to quickly predict its physical state.

[0159] The purpose of this mechanism is to enable graph neural networks to obtain reusable and generalizable physical state mapping capabilities, thereby transforming the real physical state features obtained from historical destructive detection into online lossless prediction capabilities.

[0160] In this embodiment, S5 includes the following sub-steps:

[0161] S51. Obtain the preset forming limit curve and yield strength threshold of the target material, substitute the wall thickness reduction rate and residual stress values ​​in the physical state distribution data into the preset nonlinear material failure evaluation function for joint solution, so as to obtain the risk value characterizing the margin from the failure boundary, and match the preset color mapping dictionary according to the risk value to generate a two-dimensional risk heat map.

[0162] S52. Based on spatial coordinate mapping, the pixels of the risk heat map are mapped to the corresponding vertices of the target 3D mesh model to generate a 3D multi-dimensional risk map, wherein the multi-dimensional risk map contains the risk value of each vertex;

[0163] S53. Extract the maximum value of the risk value of each vertex in the multidimensional risk map as the highest risk value;

[0164] S54. Determine the relationship between the highest risk value and the preset failure threshold;

[0165] S55. If the highest risk value is greater than or equal to the failure threshold, the evaluation decision result is output as unqualified.

[0166] S56. If the highest risk value is less than the failure threshold, the evaluation decision result is output as qualified.

[0167] This embodiment provides a risk visualization and assessment decision-making mechanism; specifically, this implementation follows the final output stage of the same exhaust bend after the physical state distribution prediction is completed, and focuses on how to convert the abstract wall thickness reduction rate and residual stress data into a visual graph, and further make a decision on whether it is qualified or not.

[0168] Specifically, while outputting only the physical state distribution data in the embodiment can provide analytical basis for R&D personnel, it still has shortcomings in actual production line applications: on the one hand, a simple numerical table is not convenient for quality inspectors to quickly locate risk locations; on the other hand, even if the system knows that there are high thinning rates and high stresses in certain locations, it may not have formed an executable release or interception decision; therefore, this embodiment introduces risk numerical calculation, heat map mapping and threshold determination mechanisms.

[0169] In step S51, the system combines the material's preset forming limit curve and yield strength threshold, and substitutes the wall thickness reduction rate and residual stress into the nonlinear material failure assessment function to calculate the risk value. The nonlinear material failure assessment function is specifically constructed as follows: Let the predicted wall thickness reduction rate of the current mesh vertex be... The predicted value of residual stress is The maximum allowable thinning rate under the primary and secondary strain states at the preset forming limit curve is [value missing]. The yield strength of the material is The system defines the risk value for this point. Let be the nonlinear joint margin function pointing from the current state point to the failure boundary in the two-dimensional state space, and its mathematical expression is:

[0170]

[0171] in, , The preset experience weighting coefficient is based on the pipe material grade; This represents an exponential function with the natural constant as its base; when the thinning rate... Approaching the limit or stress Approaching the yield threshold At that time, the joint margin weighting term within the index portion increases significantly, causing the calculated... The value rose rapidly and exponentially.

[0172] A two-dimensional risk heatmap is generated. A schematic deduction can be made: Assume the predicted results for three grid vertices are as follows: T1: thinning rate 10%, stress 120MPa; T2: thinning rate 18%, stress 210MPa; T3: thinning rate 22%, stress 250MPa. The system no longer performs simple linear weighting, but instead calculates the shortest joint margin from the current state point to the material forming limit boundary and yield boundary. If T1 is far from the failure boundary, the margin is large, and the calculated risk value is 0.32.

[0173] T2 is close to the critical boundary, with a risk value of 0.68; T3 has exceeded the critical zone of local wrinkling or cracking, with the risk value increasing dramatically to 0.84. The color mapping dictionary can be set as follows: 0 to 0.3 corresponds to blue, the safe zone; 0.3 to 0.6 corresponds to yellow, the warning zone; 0.6 to 0.8 corresponds to orange, the high-risk zone; and above 0.8 corresponds to red, the failure zone. Thus, T1 is displayed in blue, T2 in orange, and T3 in red. In this way, a two-dimensional risk heat map that truly reflects the complex stress state can be formed.

[0174] In step S52, the system maps the pixels of the two-dimensional risk heat map to the corresponding vertices of the target three-dimensional mesh model based on spatial coordinate mapping, generating a three-dimensional multi-dimensional risk map. For example, if a red pixel in the two-dimensional heat map corresponds to the outer arc area of ​​the pipe bend, it will be projected to the corresponding vertex of the outer arc of the three-dimensional pipe bend model, so that quality inspectors can intuitively see where the high-risk area is located when rotating the three-dimensional model.

[0175] In step S53, the system extracts the maximum value of the risk value of each vertex as the highest risk value; for example, the maximum value of the risk values ​​of all vertices of the current workpiece is 0.84.

[0176] In step S54, the system compares the highest risk value with the preset failure threshold; assuming the failure threshold is set to 0.80, then 0.84 has reached the danger level.

[0177] In step S55, if the highest risk value is greater than or equal to the threshold, an unqualified assessment decision result is output, and the coordinates of the high-risk vertex location, the name of the area to which it belongs, and the suggested re-inspection station can be attached at the same time; for example, the risk of the outer arc transition area of ​​the bent pipe exceeds the limit and is judged as unqualified.

[0178] In step S56, if the highest risk value is less than the threshold, a qualified evaluation decision result is output; for example, if the highest risk value is only 0.58, it can be judged as qualified and allowed to proceed to the next assembly process.

[0179] Furthermore, if the highest risk value is very close to the threshold, for example, fluctuating within a tolerance band above or below the threshold, the system can add a manual review state to avoid direct misjudgment; if the risk value of a vertex is abnormally high but exists in isolation, and its surrounding nodes are all low, then it is necessary to combine the neighborhood continuity check to determine whether it is caused by measurement noise or mapping error; if there is local occlusion or one-to-many mapping conflict when projecting the two-dimensional heat map to the three-dimensional model, then the region is updated with the average risk value or the maximum risk value of the three-dimensional grid vertices first, and the mapping confidence is recorded; if the entire workpiece does not generate a valid risk map, for example, if the preceding data is missing, then the output cannot be judged, and it guides re-collection, rather than directly making a qualified or unqualified judgment;

[0180] For example, in the final output stage of the same exhaust pipe, the system calculates a risk value of 0.84 by superimposing the high thinning rate and high residual stress in the outer arc transition area, and colors the corresponding position in the three-dimensional pipe model as red. Since this value exceeds the failure threshold of 0.80, the system automatically sorts the workpiece to the re-inspection channel and displays the unqualified evaluation result in the interface. If the highest risk value of another pipe of the same model is only 0.57, its surface is mainly yellow or light orange, the system outputs that it is qualified, and allows it to continue to flow.

[0181] The purpose of this mechanism is to transform the physical state prediction results, which are difficult to use directly, into a visible, locatable, and actionable basis for quality decision-making, so as to support online quality judgment and defect identification on the production line.

[0182] In this embodiment, the target object is a high-pressure molded part for automobiles, which includes an intake and exhaust pipe or a high-precision flange.

[0183] This embodiment provides a target object range limitation mechanism; specifically, this implementation is used to illustrate that the aforementioned method is not limited to abstract metal workpieces, but is applicable to high-pressure formed parts in automobiles, especially intake and exhaust pipes or high-precision flanges;

[0184] Specifically, if the target object is not further defined, this solution may be misunderstood as a general three-dimensional reconstruction process applicable to all ordinary mechanical parts, thus weakening the technical connection between this invention and the internal high-pressure forming process. In fact, this invention is aimed at parts with intense material flow, obvious local thinning and stress concentration during the forming process, and whose surface optical features can reflect the internal state; intake and exhaust pipes and high-precision flanges are typical representatives of this type of object.

[0185] For intake and exhaust pipes, common structures include curved sections, flared sections, flange connection sections, and local corrugated transition sections. These locations are prone to outer arc thinning, inner arc wrinkling, and stress concentration at the flange root during internal high-pressure water expansion molding. Therefore, the aforementioned curvature-texture-simulation fusion method is particularly suitable for three-dimensional reconstruction and risk assessment.

[0186] For high-precision flanges, although the overall size may be small, the flange hole perimeter, flanged area and transition fillet area have high requirements for sealing and assembly accuracy. If only geometric dimension inspection is performed, it may not be possible to find the problem of insufficient local wall thickness after flange. However, the present invention can more effectively identify potential failure areas near the flange sealing surface through joint analysis of three-dimensional geometry and surface texture.

[0187] Furthermore, if the target object is a tubular part but not obtained through internal high-pressure forming, such as a traditional welded splice, the meaning of its material flow law and surface texture characteristics may not be completely consistent with this solution. In this case, only the three-dimensional reconstruction part of this method can be used, while the physical state prediction part can be retrained or calibrated. If the target object is a variant of the same type of internal high-pressure formed part, it can also be included in the scope of application of this invention by updating the corresponding prior simulation model and training samples.

[0188] For example, on the same production line, the morning shift inspects exhaust bends that meet China VI emission standards, while the afternoon shift switches to high-precision flanges. Both types of workpieces are manufactured using an internal high-pressure forming process. The system can perform data acquisition, reconstruction, physical state prediction, and risk assessment according to the aforementioned steps, with the only difference being in the curvature-sensitive area and the prior simulation input. For exhaust bends, the focus is on the outer arc and the flange root; for high-precision flanges, the focus is on the flanged area and the periphery of the sealing surface.

[0189] The purpose of this limitation is to clarify that the object targeted by this invention has the shape-property coupling characteristics unique to internal high-pressure forming, thereby making the application boundary of the method clear and the technical effect more targeted;

[0190] In this embodiment, the prior simulation model includes a standard computer-aided design model and molding simulation data, wherein the molding simulation data includes initial water expansion pressure and initial wall thickness distribution data.

[0191] This embodiment provides a mechanism for constructing a priori simulation model; specifically, this embodiment takes over the priori input source of the same exhaust bend in steps S1 and S4 mentioned above, and is used to further illustrate that the priori simulation model is not a single model, but is composed of a standard computer-aided design model and molding simulation data.

[0192] Specifically, in this embodiment, the prior simulation model is used to assist in feature mapping and physical state prediction. If only a standard CAD model is used, the system can only obtain the theoretical geometric shape of the workpiece, but cannot know the pressure path and initial wall thickness change trend corresponding to the geometric shape during the forming process. Conversely, if only forming simulation data is used without a standard CAD model, it may be difficult to accurately complete mesh registration and position indexing. Therefore, this embodiment defines the combination of the two as a complete prior simulation model.

[0193] The standard computer-aided design model is used to provide the theoretical reference shape of the workpiece of this model; for example, for the aforementioned exhaust bend, the CAD model can clearly define its flange hole position, bending radius, straight section length and connection boundary; the target three-dimensional mesh model obtained by online reconstruction can be aligned with the CAD model to determine which position of the current mesh vertex corresponds to in the design.

[0194] The molding simulation data is used to provide process priors; for example, in the simulation of water-swelling molding, the system can obtain the initial water-swelling pressure distribution and initial wall thickness distribution data experienced by different regions during the molding process; schematically, the outer arc region of the bend may correspond to a higher material tensile state in the simulation, with an initial water-swelling pressure of 82MPa and an initial wall thickness distribution that shows a trend of transitioning from 1.50mm to 1.38mm; although this is not equivalent to the final measured value, it can provide the graph neural network with prior clues that this region is inherently more prone to thinning;

[0195] In practical use, the system can first use the CAD model to complete the coarse registration between the reconstructed mesh and the theoretical model, and then read the initial water expansion pressure and initial wall thickness value of the corresponding position in the molding simulation data according to the corresponding position index, and write it into the node features; in this way, the three-dimensional geometric features, two-dimensional texture features and process prior features form a fusion input under the same coordinate semantics.

[0196] Furthermore, if the forming simulation data of a certain type of workpiece has a low resolution, making it difficult to directly correspond to each mesh vertex, a region average mapping method can be used, that is, the process parameters of a simulation unit are assigned to all the mesh vertices it covers; if there are slight differences between the CAD model version and the current production batch, version matching and tolerance compensation can be performed before registration to prevent design changes from being misjudged as forming anomalies; if some initial wall thickness distribution data is missing, the nominal thickness in the material specification can be used as the default value, and its confidence level can be marked as lower than that of the complete simulation area;

[0197] For example, in the inspection of the same exhaust bend, the system calls the CAD model of the bend to complete the position matching of the flange end face, the bending center line and the connecting section; reads the forming simulation data and finds that the outer arc transition zone has a high water expansion pressure and a thin initial wall thickness distribution in the simulation stage; if the currently collected three-dimensional mesh vertex happens to fall in this area, the process prior will be added to its node features, so that the subsequent physical state prediction is more in line with the actual forming law;

[0198] The purpose of this limitation is to clarify that the prior simulation model simultaneously serves as both a geometric benchmark and a carrier of process knowledge, so as to improve the accuracy and interpretability of subsequent fusion predictions.

[0199] In this embodiment, the multispectral reflectance texture image is acquired by a high-resolution polarization camera, and the structured light image sequence is acquired by a laser stripe scanner.

[0200] This embodiment provides an implementation mechanism for a multimodal acquisition device. Specifically, this embodiment follows the acquisition stage of the same exhaust bend pipe at the inspection station, and is used to further illustrate the specific acquisition methods of the two types of image data in the embodiment. To ensure strict correspondence with the terms used above, in this embodiment, the structured light image sequence refers to the laser stripe image sequence continuously acquired by a laser stripe scanner under different viewpoints and different workpiece poses. The two are different descriptions of the same technical object in this specification, and no new data type is introduced. The multispectral reflective texture image refers to the set of surface reflective texture images acquired by a high-resolution polarization camera under multiple polarization angles and multiple bands.

[0201] Specifically, in the aforementioned embodiments, structured light image sequences undertake the task of geometric reconstruction, while multispectral reflectance texture images undertake the task of optical texture and material state characterization. If both types of images are acquired by ordinary industrial cameras, two problems are likely to occur in the scenario of high-pressure forming parts in metal: first, ordinary cameras have difficulty in stably acquiring high-precision depth information; second, ordinary grayscale or color images have difficulty in effectively distinguishing specular reflection, polarization reflection, and material texture information. Therefore, this embodiment limits the use of a laser stripe scanner and a high-resolution polarization camera to achieve the two types of acquisition respectively.

[0202] The laser stripe scanner projects laser stripes and acquires the stripe deformation to form a structured light image sequence. Because laser stripes have good directionality and high contour resolution, they are suitable for fine geometric scanning of the shape of bent pipes, flange boundaries, and transition fillets. Schematic, when the bent pipe rotates around the fixture, the scanner can continuously acquire stripe images from multiple angles and recover the three-dimensional coordinates of the workpiece surface. In other words, the structured light system mentioned in the previous embodiments corresponds to the laser stripe scanner in the specific hardware implementation of this embodiment, and its output is the structured light image sequence in this embodiment.

[0203] High-resolution polarization cameras are used to acquire multispectral reflectance texture images. The advantage of polarization imaging is that it can distinguish the differences in reflection in different polarization directions, thereby more sensitively capturing the specular reflection changes, texture direction changes, and micro-roughness differences of metal surfaces. If further combined with illumination or filter components of different bands, multispectral reflectance texture images can also be formed for the aforementioned reflectance change extraction and frequency domain texture analysis. Furthermore, the pixel intensity distribution information, reflectance gradient, and reflectance change data mentioned above are all calculated from this uniformly defined multispectral reflectance texture image, and their data source names do not change due to different camera operating bands or polarization angles.

[0204] In the actual workstation, the laser stripe scanner and the high-resolution polarization camera can be arranged around the same workpiece on a unified mechanical frame and establish a coordinate relationship through a common calibration plate; in this way, the three-dimensional points obtained by the scanner and the texture pixels obtained by the polarization camera can be registered in the same coordinate system, which meets the spatial coordinate registration requirements of the embodiment.

[0205] Furthermore, if the laser stripe scanner shows stripe breaks in certain highly reflective areas, this can be improved by adjusting the exposure time, changing the incident angle, or adding auxiliary anti-reflective illumination; if the polarization camera has a low signal-to-noise ratio in a certain band, this band can be skipped and the remaining effective bands can be used to complete the texture feature extraction; if there is a time synchronization deviation between the two types of equipment, such as slight vibration of the workpiece on the conveyor line, the geometric information and texture information at the same moment can be ensured by triggering the synchronizer or aligning the acquisition timestamps.

[0206] For example, after the same exhaust pipe enters the inspection station, the laser stripe scanner first acquires multiple frames of stripe images along the outer surface of the workpiece to reconstruct the three-dimensional contours of its flange end face, outer arc of the bend, and straight pipe section; the high-resolution polarization camera captures images of the surface of the bend under multiple polarization angles and multiple bands, recording its reflectivity distribution and texture direction changes; after the coordinate registration of the two types of images is completed, they can be used for the aforementioned curvature gradient calculation, reflectivity change correction, and two-dimensional texture feature extraction, respectively.

[0207] The purpose of this limitation is to clarify the hardware acquisition basis of the multimodal input of the present invention, so that both geometric information and surface optical information have sufficient sampling quality, thereby supporting subsequent three-dimensional reconstruction and physical state prediction.

[0208] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles, characterized in that, Includes the following steps: S1. Obtain the structured light image sequence and multispectral reflectance texture image of the target object, and obtain the spatial calibration parameters for multimodal data alignment. Based on the spatial calibration parameters, perform spatial coordinate registration on the structured light image sequence and multispectral reflectance texture image, and obtain the prior simulation model of the target object. S2. Calculate the local curvature gradient of a two-dimensional image based on the structured light image sequence, dynamically allocate edge extraction weights according to the local curvature gradient of the two-dimensional image, and generate a three-dimensional point cloud with non-uniform density. S3. Extract the surface reflectance variation data of the multispectral reflectance texture image, use the surface reflectance variation data as a penalty term to construct the surface reconstruction energy function, and perform three-dimensional surface fitting on the three-dimensional point cloud based on the surface reconstruction energy function to generate the target three-dimensional mesh model. S4. Extract the three-dimensional geometric features of the target three-dimensional mesh model and the two-dimensional texture features of the multispectral reflectance texture image. Construct an initial graph structure with the vertices of the target three-dimensional mesh model as graph nodes and the mesh topology edges of the target three-dimensional mesh model as graph edges. Map the three-dimensional geometric features, two-dimensional texture features and prior simulation model to the graph nodes. Calculate the feature difference between adjacent vertices and map it to the graph edges. Input the pre-trained graph neural network for feature fusion and mapping, and output the physical state distribution data representing the wall thickness reduction rate and residual stress. S5. Generate a two-dimensional risk heat map based on physical state distribution data, and overlay the risk heat map onto the surface of the target three-dimensional mesh model based on spatial coordinate mapping to generate a multi-dimensional risk map. Output the assessment and decision results of the target object based on the multi-dimensional risk map.

2. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Extract the fringe phase map from the structured light image sequence and analyze it to extract the two-dimensional image curvature data of the target object; S22. Calculate the spatial rate of change of curvature data of a two-dimensional image to generate a local curvature gradient of the two-dimensional image; S23. Determine the relationship between the local curvature gradient of a two-dimensional image and a preset curvature gradient threshold; S24. If the local curvature gradient of the two-dimensional image is greater than or equal to the preset curvature gradient threshold, then the first edge extraction weight is assigned, and a high-density point cloud region is generated based on the first edge extraction weight. S25. If the local curvature gradient of the two-dimensional image is less than the preset curvature gradient threshold, a second edge extraction weight is assigned, and a low-density point cloud region is generated based on the second edge extraction weight, wherein the first edge extraction weight is greater than the second edge extraction weight, so as to combine and generate a three-dimensional point cloud with non-uniform density.

3. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Extract pixel intensity distribution information from multispectral reflectance texture images; S32. Calculate the reflectance gradient between adjacent pixels based on pixel intensity distribution information to generate surface reflectance change data; S33. Normalize the surface reflectivity variation data to generate depth correction coefficients; S34. Embed the depth correction coefficient as a penalty term into the preset basic surface reconstruction function, wherein the basic surface reconstruction function includes a data term that constrains the three-dimensional mesh vertices to approximate the original point cloud and a smoothing term that constrains the continuity of the local surface, so as to construct the surface reconstruction energy function. S35. Run the surface reconstruction energy function to perform 3D surface fitting and spatial coordinate iterative correction on the 3D point cloud to suppress surface gloss artifacts and output the target 3D mesh model.

4. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Analyze the vertex coordinates and topological relationships of the target 3D mesh model, and compare them spatially with the standard computer-aided design model in the prior simulation model to extract the local curvature and deformation gradient of the 3D mesh, and combine the local curvature and deformation gradient of the 3D mesh into 3D geometric features. S42. Perform frequency domain transformation on the local multispectral reflectance texture image region that matches the spatial dimensions of each vertex of the target 3D mesh model to obtain frequency domain information, analyze the frequency domain information to extract the surface texture anisotropic features and surface roughness that characterize the material deformation, and combine the surface texture anisotropic features and surface roughness into the two-dimensional texture features of the corresponding graph nodes. S43. Analyze the prior simulation model to extract the initial water expansion pressure and initial wall thickness distribution data; S44. Perform multi-dimensional feature stitching on the three-dimensional geometric features, two-dimensional texture features, and initial water expansion pressure and initial wall thickness distribution data to generate the fused feature vector of the corresponding graph node; at the same time, calculate the difference between the three-dimensional geometric features, two-dimensional texture features, initial water expansion pressure and initial wall thickness distribution data between adjacent vertices to generate the edge feature vector of the corresponding graph edge. S45. Input the fused feature vector and edge feature vector into the pre-trained graph neural network for forward propagation to output physical state distribution data.

5. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, Pre-trained graph neural networks are generated through the following steps: Acquire the three-dimensional geometric features, two-dimensional texture features, and prior simulation sample data of historical samples that are the same model and use the same molding process as the target object; The actual wall thickness reduction rate and actual residual stress distribution obtained from historical samples through destructive testing are used as label data; Construct the initial graph neural network; The three-dimensional geometric sample features, two-dimensional texture sample features, and prior simulation sample data are input into the initial graph neural network for prediction to generate predicted state distribution data. Calculate the loss value between the predicted state distribution data and the label data; The network parameters of the initial graph neural network are updated using the backpropagation algorithm based on the loss value until the loss value converges, thereby generating a pre-trained graph neural network.

6. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, S5 includes the following sub-steps: S51. Obtain the preset forming limit curve and yield strength threshold of the target material, substitute the wall thickness reduction rate and residual stress values ​​in the physical state distribution data into the preset nonlinear material failure evaluation function for joint solution, so as to obtain the risk value characterizing the margin from the failure boundary, and match the preset color mapping dictionary according to the risk value to generate a two-dimensional risk heat map. S52. Based on spatial coordinate mapping, the pixels of the risk heat map are mapped to the corresponding vertices of the target 3D mesh model to generate a 3D multi-dimensional risk map, wherein the multi-dimensional risk map contains the risk value of each vertex; S53. Extract the maximum value of the risk value of each vertex in the multidimensional risk map as the highest risk value; S54. Determine the relationship between the highest risk value and the preset failure threshold; S55. If the highest risk value is greater than or equal to the failure threshold, the evaluation decision result is output as unqualified. S56. If the highest risk value is less than the failure threshold, the evaluation decision result is output as qualified.

7. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, The target is automotive internal high-pressure molded parts, including intake and exhaust pipes or high-precision flanges.

8. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, The prior simulation model includes a standard computer-aided design model and molding simulation data, in which the molding simulation data includes initial water expansion pressure and initial wall thickness distribution data.

9. The method for three-dimensional image reconstruction of high-pressure molded parts for lightweight automobiles according to claim 1, characterized in that, Multispectral reflectance texture images were acquired using a high-resolution polarization camera, while structured light image sequences were acquired using a laser stripe scanner.

Images