Image recognition-based automobile wheel cover injection molding part surface defect classification system

Abstract

The present application relates to the field of industrial vision and intelligent quality inspection, in particular to a surface defect classification system for automobile wheel cover injection molding parts based on image recognition; comprising image acquisition, two-dimensional image feature extraction, visual stress implicit mapping, spatial topology tolerance calculation and dynamic consequence classification modules; the system estimates local implicit residual stress and three-dimensional deformation probability by extracting two-dimensional defect features; the core is to map the defect to a three-dimensional CAD model, calculate the decay weight according to the topological distance to the key assembly point, and nonlinearly superimpose the basic stress risk and the deformation probability to generate an assembly failure risk index; when the index is greater than the upper limit of safety, the scrap label is output, and when the index is less than the lower limit, the release is released, and when the index is between the two, the manual reinspection is prompted; the present application realizes the transformation from only relying on the surface morphology to combining the three-dimensional physical mechanism evaluation, and can effectively avoid the misjudgment of harmless flaws and the missed detection of hidden defects.

Description

Technical Field

[0001] This invention relates to the field of industrial vision and intelligent quality inspection technology, specifically to an image recognition-based system for classifying surface defects in injection-molded automotive wheel arches. Background Technology

[0002] In the current automotive injection molding manufacturing and testing environment, three-dimensional curved surface parts such as wheel covers need to undergo surface quality inspection after they come off the production line. This is usually done by multiple industrial cameras simultaneously acquiring two-dimensional surface image data from multiple perspectives.

[0003] To determine defects in these image data, existing visual inspection solutions generally adopt a simple two-dimensional appearance classification architecture, that is, directly extracting features from the image and classifying it into planar categories based on the shape of the abnormal area. Although this solution has certain processing capabilities in conventional surface anomaly recognition scenarios, it relies heavily on two-dimensional visual appearances and is detached from the three-dimensional physical model and underlying physical mechanism of the component. It cannot accurately establish a connection between the appearance image features and the internal residual stress and local three-dimensional deformation trends. At the same time, due to the lack of mapping calculation of the topological distance from two-dimensional physical coordinates to key assembly points in three-dimensional space, the actual assembly consequences are not assessed, resulting in a high probability of misjudging and scrapping surface defects that do not pose a functional hazard, while there is a risk of missing hidden fatal defects located in critical stress areas.

[0004] Therefore, improving the accuracy and reliability of the physical mapping of multi-view two-dimensional image features to three-dimensional assembly failure risk has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an image recognition-based surface defect classification system for automotive wheel arch injection molded parts. Specifically, the technical solution of this invention includes:

[0006] The image acquisition module is used to acquire multi-view two-dimensional surface image data of the target component;

[0007] The two-dimensional image feature extraction module performs semantic segmentation on the multi-view two-dimensional surface image data to generate an initial two-dimensional defect feature set containing surface defect morphological features and two-dimensional physical coordinates.

[0008] The visual stress implicit mapping module calls a pre-set multimodal neural network model to map the initial two-dimensional defect feature set into the local implicit residual stress estimate and local three-dimensional deformation probability of the target component.

[0009] The spatial topology tolerance calculation module acquires the 3D CAD model of the target component and the coordinate data of key assembly points. Based on the global coordinate system of the inspection station, it establishes a unified 2D and 3D spatial reference. Using the 2D physical coordinates, it determines the 3D mapping position of surface defects on the 3D CAD model and assigns a local 3D deformation probability. It calculates the topological distance from the 3D mapping position to the coordinate data of the key assembly points and constructs a distance attenuation weight that decreases as the distance increases. The estimated value of implicit residual stress is multiplied by this weight to obtain the basic stress risk value, which is then input into a preset exponential decay function and nonlinearly superimposed with the local 3D deformation probability. The preset exponential decay function includes a preset convergence coefficient to control the response convergence rate under high stress conditions. The nonlinear superposition amplifies the local 3D deformation probability based on a pre-calibrated warpage amplification factor, generating an assembly failure risk index.

[0010] The dynamic consequence classification module outputs a fatal defect scrap label when the assembly failure risk index is greater than or equal to the preset safety tolerance upper limit; outputs a harmless defect release label when it is less than or equal to the preset safety tolerance lower limit; and outputs a manual re-inspection prompt label when it is between the lower and upper limits.

[0011] Optionally, the multimodal neural network model is trained based on a pre-built joint training sample set;

[0012] The joint training sample set includes injection mold flow analysis data based on computer-aided engineering and corresponding output physical image data; wherein, the two-dimensional defect feature set extracted based on the output physical image data serves as the input feature of the multimodal neural network model, and the injection mold flow analysis data serves as the supervision label of the multimodal neural network model.

[0013] Optionally, the surface defect morphology features include defect area data and texture gradient data;

[0014] Before performing semantic segmentation, the two-dimensional image feature extraction module performs Gaussian filtering smoothing on the multi-view two-dimensional surface image data to remove high-frequency environmental noise signals from the multi-view two-dimensional surface image data.

[0015] Optionally, the system further includes a process parameter fusion module;

[0016] The process parameter fusion module acquires the core process parameters in the injection molding process; wherein, the core process parameters include mold temperature data and holding pressure time data;

[0017] The visual stress implicit mapping module performs a tensor concatenation operation on the core process parameters and the initial two-dimensional defect feature set, which serves as the joint input data for the multimodal neural network model.

[0018] Optionally, the system further includes a feedback correction module;

[0019] The feedback correction module acquires actual assembly interference feedback data from the subsequent assembly stage.

[0020] The feedback correction module compares the actual assembly interference feedback data with the fatal defect scrap label or harmless defect release label output by the dynamic consequence classification module. When the comparison result is inconsistent, the feedback correction module generates a model correction instruction to drive the visual stress implicit mapping module to backpropagate and update the network weights of the multimodal neural network model. When the comparison result is consistent, the feedback correction module generates a state maintenance instruction to maintain the current network weights of the multimodal neural network model.

[0021] Optionally, the key assembly point coordinate data includes the three-dimensional coordinates of the buckle position and the three-dimensional coordinates of the locating pin position on the target component;

[0022] The topological distance is specifically the shortest geodesic distance along the three-dimensional curved surface of the target component to the three-dimensional coordinates of the buckle position or the three-dimensional coordinates of the positioning pin position.

[0023] Optionally, the two-dimensional image feature extraction module includes a multi-protocol heterogeneous data acquisition adapter;

[0024] The multi-protocol heterogeneous data acquisition adapter reads the multi-view two-dimensional surface image data through a preset industrial bus communication protocol;

[0025] The multi-protocol heterogeneous data acquisition adapter performs timestamp alignment and data format standardization processing on the multi-view two-dimensional surface image data, and uniformly maps the multi-view two-dimensional surface image data from different perspectives into multi-view two-dimensional surface image data in the global coordinate system of the detection station.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This system uses a visual stress implicit mapping module to call a multimodal neural network model, mapping the initial two-dimensional defect feature set into implicit residual stress estimates and local three-dimensional deformation probabilities. Simultaneously, the spatial topology tolerance calculation module maps two-dimensional physical coordinates to a three-dimensional CAD model, calculating the shortest geodesic topological distance from the defect location to the three-dimensional coordinates of the latch or locating pin. Based on this distance, a distance attenuation weight is constructed, and the basic stress risk value is input into an exponential attenuation function and nonlinearly superimposed with the local three-dimensional deformation probability to generate an assembly failure risk index. This mechanism overcomes the shortcomings of existing technologies that rely solely on surface morphology for classification, detached from the three-dimensional physical model and underlying physical mechanisms. It accurately assesses the real impact of defects on actual assembly, effectively avoiding the misjudgment and scrapping of non-functional hazardous defects and the missed detection of latent defects in key stress areas, significantly improving the reliability of the judgment.

[0028] 2. This system pre-constructs a joint training sample set, uses injection mold flow analysis data based on computer-aided engineering as supervision labels, and trains a multimodal neural network model by combining it with a two-dimensional defect feature set extracted from the output physical object image data. This establishes a reliable physical mapping between surface visual features and internal residual stress and deformation trends. In addition, the system sets up a process parameter fusion module to acquire core process parameters such as mold temperature data and holding time data, and performs tensor splicing operation with the initial two-dimensional defect feature set as joint input. This multi-dimensional fusion mechanism enables the model to accurately correct visual features by combining molding process background information, effectively distinguishing samples with similar appearances but different physical consequences under different processes, further enhancing the accuracy of model inference.

[0029] 3. This system is equipped with a multi-protocol heterogeneous data acquisition adapter in the 2D image feature extraction module. After reading the image through the preset industrial bus communication protocol, it performs timestamp alignment and data format standardization processing to uniformly map the image data from different perspectives to the data in the global coordinate system of the inspection station. This solves the problem of data unification under multiple devices and multiple perspectives, and provides a reliable coordinate basis for accurate 3D mapping. In addition, before performing semantic segmentation to obtain the surface defect morphology features containing defect area data and texture gradient data, the system performs Gaussian filtering smoothing operation on the multi-view 2D surface image data. This effectively removes high-frequency environmental noise signals and avoids erroneous extraction caused by environmental reflection or discrete disturbances, ensuring that defect morphology features with process significance can be captured clearly and stably. Attached Figure Description

[0030] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0031] 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.

[0032] like Figure 1 As shown, an image recognition-based surface defect classification system for automotive wheel arch injection molded parts includes,

[0033] The image acquisition module is used to acquire multi-view two-dimensional surface image data of the target component;

[0034] The two-dimensional image feature extraction module performs semantic segmentation on multi-view two-dimensional surface image data to generate an initial two-dimensional defect feature set containing surface defect morphological features and two-dimensional physical coordinates.

[0035] The visual stress implicit mapping module calls a pre-built multimodal neural network model to map the initial two-dimensional defect feature set into the local implicit residual stress estimate and local three-dimensional deformation probability of the target component.

[0036] The spatial topology tolerance calculation module acquires the 3D CAD model of the target component and the coordinate data of key assembly points. Based on the global coordinate system of the inspection station, it establishes a unified 2D and 3D spatial benchmark. It then determines the 3D mapping position of surface defects on the 3D CAD model based on the 2D physical coordinates and assigns a local 3D deformation probability. The module calculates the topological distance from this 3D mapping position to the coordinate data of the key assembly points and constructs a distance attenuation weight that decreases as this distance increases. The estimated implicit residual stress is multiplied by this weight to obtain the basic stress risk value, which is then input into a preset exponential attenuation function and nonlinearly superimposed with the local 3D deformation probability. The preset exponential attenuation function includes a preset convergence coefficient to control the response convergence rate under high stress conditions. The nonlinear superposition amplifies the local 3D deformation probability based on a pre-calibrated warpage amplification factor, generating an assembly failure risk index.

[0037] The dynamic consequence classification module outputs a fatal defect scrap label when the assembly failure risk index is greater than or equal to the preset safety tolerance upper limit; outputs a harmless defect release label when it is less than or equal to the preset safety tolerance lower limit; and outputs a manual re-inspection prompt label when it is between the lower and upper limits.

[0038] This embodiment provides a surface defect consequence classification mechanism for online quality inspection and subsequent final assembly collaborative judgment of automotive wheel arch injection molded parts. Specifically, the system is deployed between the off-line inspection station and the pre-assembly buffer station in the wheel arch injection molding workshop, and the inspection object is the left front wheel arch injection molded part of the same vehicle model. This wheel arch needs to cooperate with the buckles and positioning pins in the body liner area during subsequent assembly, so not all surface abnormalities will cause assembly risks. On the contrary, some dark spots with contrast below the preset grayscale threshold and flow marks with a size smaller than the preset area parameter, if distributed within the preset distance range of the high-stress assembly area, may induce buckle cracking, hole displacement, or interference with adjacent sheet metal parts during press-fitting. This embodiment uses whether it affects the continuity of subsequent assembly and structural reliability as the judgment target, rather than simply classifying the defect name in a plane.

[0039] Specifically, the image acquisition module can consist of multiple industrial cameras, directional light sources, and workpiece positioning fixtures set around the inspection station. When the wheel cover injection molded part is in place, the fixture fixes it in a preset posture, and the cameras simultaneously acquire two-dimensional surface images from multiple perspectives, such as the outer surface, inner surface, side edge, and buckle perimeter. The reason for adopting multi-view acquisition is that the wheel cover is a three-dimensional curved surface part, and shadow occlusion, edge distortion, or local reflection are easy to occur under a single perspective, making it difficult to fully express certain flow marks and hole edge shrinkage extending along the ribs in a single image.

[0040] After receiving the above multi-view images, the two-dimensional image feature extraction module performs semantic segmentation on the defect regions in the images to obtain an initial two-dimensional defect feature set. The semantic segmentation operation uses a fully convolutional neural network based on an encoder-decoder architecture to perform binary or multi-class classification on the image pixels, and outputs a pixel-level mask matrix containing the precise topological boundaries of the defects. The system calculates the number of connected pixels in the mask matrix through an algorithm, and combines it with camera calibration parameters to convert it into defect area data. At the same time, it performs derivative calculation on the mask boundaries to extract texture gradient data.

[0041] This feature set contains at least two types of information: first, the morphological information of surface defects, such as defect boundaries, extension direction, area range, and brightness variation characteristics; second, the two-dimensional physical coordinates of the defects at the inspection station. These two-dimensional physical coordinates are not simply pixel positions, but planar positions that have been calibrated and correspond to the actual dimensions of the workpiece, enabling subsequent modules to correlate anomalies on the image with structural regions of the workpiece entity. For components like wheel covers, this step is significant in converting apparent image anomalies into physical location descriptions that can be integrated with three-dimensional engineering models.

[0042] The visual stress implicit mapping module calls a pre-built multimodal neural network model to infer the initial two-dimensional defect feature set and output the local implicit residual stress estimate and the local three-dimensional deformation probability. The multimodal neural network model adopts a dual-branch fusion network architecture, specifically including: a convolutional neural network feature extraction branch for processing the two-dimensional defect feature set and image data, and a multilayer perceptron encoding branch for processing structured features such as core process parameters. The feature vectors extracted by the two branches are fused through a tensor concatenation layer and then input to a fully connected prediction layer, finally outputting continuous local implicit residual stress estimates and local three-dimensional deformation probabilities.

[0043] The output is not directly measured by the camera, but is an implicit physical inference obtained based on long-term accumulated production samples, mold flow analysis results, and physical comparison relationships. Its industrial mechanism is that the flow marks, air bubbles, shrinkage, bright spots, and weld lines on the surface of injection molded parts are essentially physical morphological mappings of uneven shrinkage during melt filling, holding pressure and compensating for shrinkage, and cooling on the surface. These surface features are often stably correlated with internal chain segment orientation, uneven volume shrinkage, and residual stress concentration. Therefore, this module not only identifies the appearance characteristics of defects, but also further outputs physical inferences of the corresponding local stress being too high or the warping tendency being enhanced.

[0044] The spatial topology tolerance calculation module reads the three-dimensional computer-aided design model data of the wheel arch and the coordinate data of key assembly points, and maps the two-dimensional physical coordinates to the corresponding area on the three-dimensional model. For the same surface anomaly, if it is located in the inner area that is not load-bearing, not mating, and not visible after installation, its impact on the assembly function is lower than the preset tolerance. However, if it is located at the root of the buckle, the edge of the mounting hole, or the flange area adjacent to the body sheet metal, the assembly failure risk index will exceed the preset safety tolerance.

[0045] To this end, the module calculates the topological distance from the three-dimensional mapping location of the defect to the key assembly point and establishes attenuation weights based on the distance. The topological distance is used here instead of the straight-line distance because the wheel cover is a continuous curved surface. Assembly stress and deformation propagation often pass along the curved surface of the shell and the rib path. Straight-line penetrating measurement cannot truly reflect the force coupling relationship.

[0046] This module combines the implicit residual stress estimate with the distance attenuation weight to obtain the basic stress risk value, and then nonlinearly superimposes it with the local three-dimensional deformation probability to form the assembly failure risk index. The so-called nonlinear superposition means that when there is an internal stress greater than the preset stress threshold and a warping tendency greater than the preset deformation probability threshold in a local area, the risk growth is not a simple addition, but a nonlinearly increasing coupled amplification effect. This is because high stress will reduce the material's bearing margin during press fitting, and warping will reduce the geometric assembly margin. The combined effect of the two makes it easier to induce fracture or interference.

[0047] Specifically, to avoid excessively high local foundation stress risk values ​​leading to an uncontrolled overflow of the final risk index, this embodiment employs an upper bound convergence mechanism for the preset exponential decay function. This means that as the input foundation stress risk value increases, the growth rate of its output risk contribution decreases exponentially. For example, the preset exponential decay function is configured as follows: Based on the foundation stress risk value Preset convergence coefficients and natural constants Calculate the converged stress risk response term ;

[0048] Perform a nonlinear superposition operation on the response term. With local three-dimensional deformation probability Nonlinear composite amplification can be achieved using the formula... Combined with warp amplification factor Calculate the final assembly failure risk index ;in, Represents the probability of local three-dimensional deformation. Squaring operations are performed to reflect the high-risk nonlinear amplification characteristics; through the above quantification rules, the physical logic of high-stress convergence and the mutual amplification of composite deformation is clearly defined;

[0049] In order to ensure the numerical stability of the nonlinear superposition formula, the foundation stress risk value is... With local three-dimensional deformation probability All values ​​were normalized before being entered into the formula. Interval; preset convergence coefficient The value range is set to Used to control the exponential convergence rate under high stress conditions to prevent response overflow; warpage amplification factor Engineering calibration is then performed based on the flexural modulus and local wall thickness of the material used in the target injection molded part, with the value range set as follows: ;

[0050] The dynamic consequence classification module performs classification arbitration on the assembly failure risk index; the system pre-sets a preset upper limit and a preset lower limit for safety tolerance, ensuring that the upper limit is higher than the lower limit; the upper and lower limits for risk safety tolerance are not randomly assigned, but are determined based on the statistical rate of buckle breakage and press-fit interference deviation records of historical batches of this vehicle model at the final assembly station, combined with maximum likelihood estimation fitting; or they are obtained through stress limit calibration by pre-constructed computer-aided engineering interference collision simulation of whole vehicle assembly.

[0051] If the assembly failure risk index is higher than the upper limit, the probability of the part breaking during assembly or having a positioning deviation is higher than the preset scrap threshold, which may lead to positioning deviation or frictional interference with adjacent parts. Therefore, a fatal defect scrap label will be output.

[0052] If the assembly failure risk index is below the lower limit, the surface abnormality is considered not to constitute a functional hazard, and a harmless defect release label is output. If the assembly failure risk index falls between the two, it indicates that the uncertainty is high, and a manual re-inspection prompt label is output. Quality inspectors then conduct further confirmation by combining supplementary visual inspection, trial assembly, or local measurement. This hierarchical logic prevents the system from rejecting all visual abnormalities and also avoids simply releasing hidden high-risk parts.

[0053] As an anomaly handling mechanism, when there are missing viewpoints, local overexposure, workpiece positioning misalignment, or surface contamination during image acquisition, the system can mark the area as a low-confidence area. For low-confidence areas, on the one hand, it can trigger a retake or switch to a backup light source; on the other hand, if it is not possible to retake the image in time, the judgment result corresponding to the area can be directly sent to the manual re-inspection channel to avoid outputting a definitive scrap or release conclusion when the input basis is unreliable.

[0054] Furthermore, if the 3D model version is inconsistent with the current production batch workpiece version, or if the coordinates of key assembly points cannot be called, the spatial topology tolerance calculation module will pause the final generation of the assembly failure risk index. At this time, the system can revert to outputting only conservative labels indicating the presence of surface anomalies that require manual confirmation, to prevent incorrect release due to model mismatch.

[0055] During a continuous production shift of a certain model's wheel cover, after multi-view acquisition, a dark spot-like area with a smaller area than a preset area threshold but continuous texture was identified near the upper snap rib root of a left front wheel cover. From a purely visual perspective, this area is not obvious and may be classified as a minor flow mark if classified according to traditional defect names. However, in this embodiment, after mapping, this area was found to be located in the area adjacent to the curved surface of the snap rib root. The implicit mapping result of visual stress shows that there are signs of high residual stress at this location and a certain probability of warping.

[0056] Spatial topology tolerance calculation further determined that the distance from the nearest snap-fit ​​position along the curved surface was less than the preset safety distance threshold. Therefore, the assembly failure risk index was raised to above the upper limit of the safety tolerance, and the system output a fatal defect scrap label. In contrast, another wheel cover in the same shift had a thin scratch on the inner non-visible surface far from the assembly hole. Although semantic segmentation could identify it stably, the distance between its three-dimensional mapping position and the key assembly point was greater than the preset safety distance threshold, and it did not show obvious residual stress and warping tendency. It was finally judged as a harmless defect release label.

[0057] The purpose of this step is to classify surface image anomalies in conjunction with internal stress state, three-dimensional deformation trend, and assembly topology, thereby distinguishing between visible defects and functionally hazardous defects, improving the accuracy of defect detection, and ensuring the structural reliability of subsequent assembly.

[0058] In this embodiment, the multimodal neural network model is trained based on a pre-built joint training sample set;

[0059] The joint training sample set includes injection mold flow analysis data based on computer-aided engineering and corresponding output physical image data; among them, the two-dimensional defect feature set extracted based on the output physical image data serves as the input features of the multimodal neural network model, and the injection mold flow analysis data serves as the supervision label of the multimodal neural network model.

[0060] This embodiment provides a joint training mechanism for establishing the correspondence between visual features and internal physical states. Specifically, in the aforementioned online detection scenario, if only manual experience is relied upon to label images with names such as flow marks, shrinkage, and scratches, although a conventional appearance recognition model can be trained, it is difficult to stably infer residual stress and warpage risk. This is because the same visual defect has different internal physical consequences under different gate layouts, wall thickness distributions, and cooling conditions. Therefore, this embodiment further introduces injection mold flow analysis data based on computer-aided engineering as a source of supervision during the model training phase.

[0061] Specifically, the joint training sample set consists of two paired data parts: one part is the physical image data of each batch or each type of typical process condition, and the other part is the mold flow analysis data corresponding to the physical object; the mold flow analysis data may include the pressure distribution at the filler end, temperature field distribution, volume shrinkage trend, weld line location, potential residual stress distribution, and warpage prediction area, etc.; the physical images are preprocessed and feature extracted to form a two-dimensional defect feature set, which is used as the model input; the stress and deformation information of the corresponding area in the mold flow analysis is used as supervision labels; this training method aims to establish the mapping relationship between surface appearance and internal physical consequences, rather than simply classifying apparent defects;

[0062] In a simplified verification model, the surface of the wheel cover of a batch of samples can be divided into several local regions. For example, region Z1 is located at the root of the buckle, region Z2 is located on the inner non-load-bearing surface, and region Z3 is located near the flange of the mounting hole. After extracting the physical image, Z1 shows fine flow marks and slight dark spots, Z2 shows scratches with grayscale contrast greater than a preset threshold, and Z3 shows small-area shrinkage. The corresponding mold flow analysis results show that the residual stress of Z1 exceeds the preset reference stress value, the stress level of Z2 is lower than the reference stress value, and the probability of local three-dimensional deformation of Z3 is greater than the preset lower probability limit. The model training establishes a mapping relationship from appearance features to internal physical features to achieve the following goal: based on the two-dimensional morphology and position information provided by the image side, combined with the corresponding molding mechanism physical quantities provided by the mold flow analysis side, the pairing of the two enables the model to infer the internal state from the visual appearance.

[0063] Mold flow analysis is used as a monitoring label because residual stress, orientation inhomogeneity, and shrinkage differences during injection molding are often difficult to measure directly on each full-volume physical part. Computer-aided engineering can provide relatively stable field distribution predictions under uniform material, mold, and process boundaries. Although mold flow analysis is not the final measured value, its indication of high-risk areas has engineering reference significance and can be used as a physical prior during the training phase. Through this training method, when the system is running online, even when facing ordinary production parts that have not been directly sectioned or stress tested, it can output physically meaningful estimation results based on image features.

[0064] As an anomaly handling mechanism, if a batch of physical images lacks process consistency with the mold flow analysis samples—for example, the mold version has changed, the gate position has been adjusted, the material grade has been replaced, or the wall thickness design has been modified—then this batch of data will not be directly included in the training set to prevent the model from learning incorrect correspondences. For such samples, they can first be entered into the verification sample pool, and it will be decided whether to use them for incremental training after confirmation by the process engineer. For another example, if the physical image has severe contamination, missing corners, or missing images, making it impossible to reliably extract local features, then this sample can only retain the mold flow analysis data as a reference and will not participate in the strongly supervised training from image to label.

[0065] In the initial stage of the left front wheel cover project for the same model, the process team generated multiple sets of mold flow analysis results under different mold temperatures, holding pressures, and cooling conditions during the mold trial stage. At the same time, they acquired multi-view images of the corresponding physical wheel cover on the trial production line. For local areas near the locating pin holes, if the image continuously shows appearance features such as gray values ​​lower than the preset background gray value and blunted edge transitions, and the mold flow analysis continuously indicates that there is a shrinkage gradient and warping trend in this area that is greater than the preset gradient threshold, then after multiple rounds of training, the model will establish a stable correspondence between these visual features and high assembly risk. Conversely, for scratches that are far from the assembly area and have a depth less than the preset depth threshold, even if the contrast in the image is high, the model will not incorrectly map them as high-risk signals because the mold flow analysis does not show that they are accompanied by stress concentration.

[0066] The purpose of this mechanism is to provide an engineering-interpretable training basis for implicit mapping of visual stress, thereby enabling reliable inference from two-dimensional image features to residual stress estimates and local three-dimensional deformation probabilities, and improving the system's responsiveness to real assembly consequences.

[0067] In this embodiment, the surface defect morphology features include defect area data and texture gradient data;

[0068] Before performing semantic segmentation, the two-dimensional image feature extraction module performs Gaussian filtering smoothing on the multi-view two-dimensional surface image data to remove high-frequency environmental noise signals in the multi-view two-dimensional surface image data.

[0069] This embodiment provides a feature extraction mechanism for noise reduction and morphological enhancement of multi-view surface images. Specifically, in the aforementioned online detection scenario, relying solely on the original image for semantic segmentation can easily misidentify environmental reflections, texture particles, micro-processing marks on the mold surface, and edge burrs caused by conveyor vibrations as real defects. In particular, the surface of wheel cover injection molded parts is mostly curved and transitioning between ribs. When the incident angle of illumination deviates beyond the preset angle tolerance, high-frequency bright spots or stripes may be formed. If smoothing is not performed first, the subsequent segmentation results will show fragmentation, boundary jumps, or false patches.

[0070] Specifically, the 2D image feature extraction module performs Gaussian filtering smoothing on multi-view images before segmentation. The engineering significance of this operation is to suppress high-frequency environmental noise and preserve the main trend of the defect contour. The actual forming defects on the wheel cover surface are usually manifested as areas of continuous variation in a certain area. For example, shrinkage pits will form gradually changing shadows, flow marks will appear as continuous textures extending along the flow direction, and weld lines will mostly appear as narrow and long but directionally stable band-shaped areas. Environmental noise often manifests as isolated sharp pixels, local sudden reflections, or random particles. After smoothing, the system can more effectively separate the continuous morphology related to the process mechanism from the discrete disturbances related to the shooting environment.

[0071] Regarding defect morphology characteristics, this embodiment extracts at least defect area data and texture gradient data. Defect area data reflects the coverage of abnormal areas, which helps to distinguish between local minor scratches and large-scale molding unevenness. Texture gradient data reflects the steepness of surface grayscale or texture changes, and can be used to characterize whether the defect edges are abrupt or have flow directionality. For injection molded parts, large-area dark spots with a texture gradient change rate lower than a preset threshold are more likely to correspond to uneven internal cooling or shrinkage differences. On the other hand, thin and elongated texture gradients with clear directions are closer to melt flow traces. Through the combination of these two types of features, the system can provide more process-meaning inputs for the subsequent physical mapping module.

[0072] In a simplified verification scenario, the original segmentation candidate region in an image from a certain perspective may contain three adjacent patches A, B, and C, where A and B are real continuous flow trace regions, and C is a single strong reflective noise point. After Gaussian filtering, the continuous trend between A and B is preserved, while the isolated peak of C is weakened. Semantic segmentation is more likely to identify A and B as the same defect zone, while excluding C from the valid defects. At this time, the extracted area data is closer to the real defect range, and the texture gradient direction is more consistent with the physical texture formed by melt flow.

[0073] As an anomaly handling mechanism, if the wheel cover surface itself has a high-gloss coating, mold release agent residue, or local water vapor condensation, simple smoothing may still retain some false features. In this case, the system can combine multi-view consistency for secondary verification, that is, only when the same physical area shows continuous anomalies in adjacent views will it be included in the high-confidence defect set. If the image details are excessively smoothed after Gaussian filtering, resulting in blurred boundaries of crack-like defects with pixel widths lower than the preset value, the system can use a smoothing intensity lower than the preset filter kernel size for key areas such as the buckle root and hole edge, or retain the original image branches in parallel to prevent real dangerous defects from being eliminated.

[0074] In the same wheel cover inspection station, dense bright spots appeared near the inner rib of a wheel cover due to changes in the illumination angle. In the original image, these bright spots were mixed with slightly shrunken areas. After smoothing, the random bright spots were significantly reduced, while the continuous shadow contours formed by the shrunken areas were still preserved. The system extracted the small but complete area features of this region, as well as the texture gradient features that gradually transitioned from the center to the edge, and sent them to the subsequent implicit mapping module. Compared with the case without noise reduction, the segmentation results were no longer interfered with by a large amount of fragment noise.

[0075] The purpose of this step is to achieve a stable representation of real surface defects by first suppressing high-frequency environmental noise and then extracting area and texture gradient features with process significance, thereby providing a clean and interpretable input basis for subsequent residual stress and deformation risk inference.

[0076] In this embodiment, the system further includes a process parameter fusion module;

[0077] The process parameter fusion module acquires the core process parameters in the injection molding process; among them, the core process parameters include mold temperature data and holding pressure time data;

[0078] The visual stress implicit mapping module performs tensor splicing operations on the core process parameters and the initial two-dimensional defect feature set, which are then used as joint input data for the multimodal neural network model.

[0079] This embodiment provides a fusion mechanism that combines injection molding process parameters with visual defect features. Specifically, in the aforementioned scheme that infers internal stress and deformation risk solely based on image appearance, it can already cover most stable production scenarios. However, under conditions deviating from the standard process parameter range, different molding batches may exhibit similar appearances but different internal risks. This is because the residual stress and warpage of wheel cover injection molded parts are not entirely determined by surface images alone, but are also closely related to process conditions such as mold temperature and holding time.

[0080] For example, the same flow mark, when the mold temperature is too low and the holding pressure is insufficient, often means poor fusion of the melt front and insufficient feeding. The corresponding internal stress and dimensional stability risks may be higher than those of the same appearance feature formed under normal process. Therefore, this embodiment introduces a process parameter fusion module to perform process background correction on visual judgment.

[0081] Specifically, the process parameter fusion module reads the core process parameters corresponding to the current wheel cover part from the injection molding machine, mold temperature controller, or workshop manufacturing execution system. These parameters include at least mold temperature data and holding time data. Mold temperature determines the cooling rate of the melt in the cavity and the surface freezing behavior. When it is below the preset mold temperature lower limit, it is easy to form discontinuous flow at the filling end, poor welding, and high freezing stress. When it is above the preset mold temperature upper limit, it may affect the shrinkage balance after demolding. Holding time determines whether the shrinkage compensation can be fully carried out. When it is shorter than the preset holding time threshold, the thick wall and hole edge areas are more prone to shrinkage and uneven volume shrinkage, causing local warping. These process variables are input into the model together with the initial two-dimensional defect feature set, so that the model can make joint inferences by combining the molding process background information while extracting surface visual features.

[0082] In this embodiment, tensor splicing can be understood as organizing data from different sources into a joint input structure according to a unified sample number. For example, the image branch of a wheel cover outputs a local defect feature vector F, which contains information such as area, texture gradient, and two-dimensional coordinates; the process branch outputs a parameter vector C, which contains the mold temperature state and holding time state corresponding to the part; the model receives the combined joint vector, rather than two isolated sets of information; in this way, the system realizes the regular mapping of the same appearance corresponding to different physical consequences under different processes.

[0083] In a simplified application scenario, two wheel covers can be designated as samples D1 and D2. Both exhibit similar slight flow marks in the image near the buckle, but D1 corresponds to a mold temperature within a stable window and sufficient holding pressure, while D2 corresponds to a low mold temperature and shortened holding pressure. After joint input, the model output will characterize the implicit residual stress estimate for D2 as a risky part that is more likely to be accompanied by insufficient feeding and high internal stress, while giving a lower risk estimate for D1. This difference does not come from subjective human rules, but from the statistical learning results under the combined effect of process and image.

[0084] As an anomaly handling mechanism, if a wheel cover is missing a corresponding process parameter, such as due to equipment communication interruption, batch binding failure, or data acquisition delay, the system can adopt a two-level processing approach. First, if the missing parameter is only a single parameter and can be estimated from the same stable interval, the process reference value within the most recent stable production window can be used as a temporary supplement, while a low confidence mark is added to the result. Second, if the key parameters are missing as a whole or the risk of batch mismatch is high, the visual stress implicit mapping module reverts to the basic mode that relies solely on image features, and prioritizes sending boundary samples for manual re-inspection to avoid making a release conclusion with low reliability when the process background is unknown.

[0085] During the continuous production of wheel covers for the same model, the mold temperature fluctuated briefly after the start of the night shift, causing the mold temperature to deviate from the normal range for a period of time. During this period, if only images were viewed, many wheel covers only showed slight dark lines near the locating pins, making it difficult to distinguish the level of risk. After introducing process parameters, the system identified that these dark lines appeared under the process background of low mold temperature and short holding pressure. Therefore, the corresponding estimated value of implicit residual stress was raised as a whole. Some samples near the locating pin flange were classified as manual re-inspection or scrapped, while similar appearance samples formed under the stable process of the day shift were allowed to pass.

[0086] The purpose of this mechanism is to provide the forming process context for visual features, thereby enabling more precise inferences about internal stress and deformation risks and reducing misjudgments caused by relying solely on appearance.

[0087] In this embodiment, the system further includes a feedback correction module;

[0088] The feedback correction module acquires actual assembly interference feedback data from subsequent assembly stages;

[0089] The feedback correction module compares the actual assembly interference feedback data with the fatal defect scrap label or harmless defect release label output by the dynamic consequence classification module. When the comparison result is inconsistent, the feedback correction module generates a model correction instruction, which drives the visual stress implicit mapping module to backpropagate and update the network weights of the multimodal neural network model. When the comparison result is consistent, the feedback correction module generates a state maintenance instruction to maintain the current network weights of the multimodal neural network model.

[0090] This embodiment provides a feedback correction mechanism that combines the actual assembly results of the subsequent process to perform closed-loop correction of the risk model of the preceding process. Specifically, in the aforementioned scheme, the system can already make scrap, release, or re-inspection judgments for wheel cover parts based on images, processes, and 3D topology information. However, both the injection molding process and the final assembly environment have long-term drift characteristics. For example, mold wear, material batch differences, fixture micro-offsets, and changes in buckle stiffness will all cause the correspondence between visual features, physical state, and assembly consequences to change over time. If the model is not corrected for a long time, it may perform well in the early stages but gradually deviate from the actual assembly results in subsequent production. Therefore, this embodiment further introduces a feedback correction module.

[0091] Specifically, the feedback correction module acquires actual assembly interference feedback data from the subsequent assembly stage; this feedback data can come from abnormal pressing force records, buckle breakage records, locating pin insertion difficulty records, defect codes from the rework station, or interference detection results from offline trial assembly tooling.

[0092] The module compares these subsequent results with the critical defect scrap label or harmless defect release label given by the preceding dynamic classification module. If the preceding process determines that the sample is releaseable, but the subsequent process shows obvious assembly interference, buckle damage, or hole misalignment, it indicates that the preceding model has missed a judgment. If the preceding process determines that the sample is scrapped, but the control sample does not show interference after subsequent trial assembly, it indicates that the preceding model has an overly conservative tendency. In case of inconsistency, the feedback correction module generates a model correction instruction to drive the visual stress implicit mapping module to update the model weights. When consistent, the current parameters are maintained to ensure stable operation.

[0093] In a simplified verification model, three samples E1, E2, and E3 are found in a batch of wheel cover parts. The front-end system outputs "release," "scrap," and "release" respectively. The back-end assembly feedback shows that E1 was successfully pressed in, E2's trial assembly interference was lower than the preset assembly tolerance limit, and E3 developed root cracks during the snap-fit ​​pressing stage. Therefore, E1 is consistent with the front-end judgment, while E2 and E3 are inconsistent. The feedback correction module sends E2 as an over-judged sample and E3 as an under-judged sample back to the model correction process. This increases the response weight of the model parameters to high-risk phenomena like E3 in subsequent iterations, while reducing the penalty weight for features like E2 that have no actual consequences. In this correction mechanism, since the actual assembly interference feedback data in the back-end is represented by discrete states such as cracking, interference, and normal operation, while the model prediction values ​​are continuous implicit residual stress and deformation probability, there is a difference in the evaluation domain. Therefore, this embodiment establishes the calculation basis for backpropagation through pseudo-label offset compensation.

[0094] Specifically, when a missed detection occurs, such as E3 (where measured interference is not met but the decision to allow is not), the feedback correction module adds a preset positive penalty bias to the current model's output stress estimate and deformation probability to generate a corrected target pseudo-label, which is then used as the true label to calculate the mean squared error loss function. When an over-detection occurs, such as E2, a negative tolerance bias is added to generate a target pseudo-label. After calculating the error gradient, the backpropagation algorithm is called to adjust the parameters of the multimodal neural network. This structured mapping effectively establishes a computational mechanism for inferring continuous physical quantities from discrete assembly feedback and then fine-tuning it in the reverse direction.

[0095] The technical advantage of this closed-loop correction lies in using the dimensional fit and structural stability of the actual assembly as the benchmark for model optimization; by establishing a back-end feedback mechanism for the automotive parts inspection system, the model's judgment results can be continuously aligned with the actual physical verification data at the final assembly site, thereby improving the reliability of the inspection system in the actual production environment.

[0096] As an anomaly handling mechanism, if the feedback data from the downstream process is incomplete, such as some released parts not being traced to the final assembly result, or assembly anomalies caused by the superposition of tolerances of other body parts and not simply attributed to the wheel arch part itself, then such samples will not directly participate in strong correction, but will first enter the identification buffer pool; only when the part identity is traceable, the cause of the anomaly is attributable, and the feedback record is complete, will the formal model correction be triggered; furthermore, when concentrated feedback anomalies occur in a short period of time, the system can also set up an update throttling mechanism to avoid abnormal fluctuations in model parameters caused by occasional external interference in a single shift; during this period, priority is given to maintaining the stability of the current model, while notifying process personnel to check the status of equipment and assembly fixtures;

[0097] Three months after the aforementioned left front wheel cover project entered mass production, several wheel covers experienced insertion forces exceeding the preset assembly force limit during the upper buckle pressing process within a certain week. Tracing back to previous records revealed that most of these samples were judged by the system as harmless defects and released, and their image features were concentrated in dark lines with grayscale contrast below the preset value near the buckle. After the feedback correction module established a correspondence between these samples and the actual interference records, it generated model correction instructions to improve the model's risk sensitivity to such dark lines with grayscale contrast below the preset value in specific locations. After subsequent updates, the system output more manual re-inspection or scrap labels for similar samples, while continuing to release samples with inner non-load-bearing scratches that had not been verified to have problems in subsequent processes.

[0098] The purpose of this mechanism is to continuously calibrate the risk inference of the preceding stage using the actual assembly results, so as to update the model as the production status evolves and improve the consistency between the judgment results and the actual assembly consequences.

[0099] In this embodiment, the key assembly point coordinate data includes the three-dimensional coordinates of the buckle position and the three-dimensional coordinates of the locating pin position on the target component;

[0100] The topological distance is specifically the shortest geodesic distance along the three-dimensional curved surface of the target component to the three-dimensional coordinates of the buckle position or the three-dimensional coordinates of the positioning pin position.

[0101] This embodiment provides a mechanism for defining key assembly points and calculating geodesic distances on the curved surface of a wheel arch. Specifically, in the aforementioned spatial topology tolerance calculation, if only a single spatial straight-line dimension is used as the basis for risk attenuation based on the spatial straight-line distance from the defect to the assembly point, although it is simple to implement, it is prone to distortion for wheel arches with flanges, arches, ribs, and hole transitions. This is because the force propagation and assembly deformation of the wheel arch do not travel directly through the air to another point, but are gradually transmitted along the component shell, ribs, and local curved surfaces. Therefore, this embodiment specifically defines the key assembly points as the three-dimensional coordinates of the buckle position and the three-dimensional coordinates of the locating pin position, and defines the distance from the defect to these points as the shortest geodesic distance along the three-dimensional curved surface of the wheel arch to that point.

[0102] Specifically, the snap-fit ​​position usually corresponds to the area of ​​concentrated pressure during press fitting, while the locating pin position corresponds to the assembly reference constraint area. The former focuses more on local brittle fracture and snap-fit ​​failure, while the latter focuses more on hole position offset and overall assembly posture deviation. After the system reads the 3D CAD model, it can predefine several sets of key assembly points on the model, such as the upper edge snap-fit ​​point set, the side snap-fit ​​point set, and the locating pin hole center point set. For each detected surface anomaly, the system first finds its mapping position on the 3D model, and then searches along the model surface for the shortest path length to the nearest key assembly point, using it as the topological distance. The shorter the distance, the closer the defect is to the key pressure or positioning area, and the more direct its potential impact.

[0103] The shortest geodesic distance is used instead of Euclidean distance because wheel covers often have large curved surfaces and folded edges. Although the straight-line distance between a defect point and the positioning pin hole in three-dimensional space is less than the first preset threshold, if there is a flange or rib transition between them, the actual stress transmission path and assembly influence path will be longer. Conversely, defects distributed along the same continuous curved surface, even if the straight-line distance is greater than the first preset threshold, may still have a direct impact on the buckle root through the shell connectivity. The geodesic distance is more consistent with the continuous transmission mechanism in actual structures.

[0104] In a simplified verification scenario, we can assume that there are defect points Q1 and Q2 on the wheel cover surface, and the nearest critical assembly point is the snap-fit ​​point K. Looking at it in terms of spatial straight lines, the straight-line distance between Q1 and K is less than the straight-line distance between Q2 and K. However, Q1 is located on the back of the flange and needs to cross a geometric broken line to connect with the surface where K is located, while Q2 is located on the same continuous arc surface as K. After searching along the surface, the shortest geodesic path from Q2 to K is actually shorter. Therefore, Q2 has a higher risk weight in this embodiment. This processing mechanism better reflects the actual physical transmission characteristics of deformation along the shell and force transmission along the surface during final assembly.

[0105] As an exception handling mechanism, if some old CAD data lacks complete surface definitions and cannot directly obtain stable geodesic paths, the system can degenerate into approximate surface distance calculation based on discrete grids. That is, the wheel cover surface is discretized into grid nodes, and the shortest path of the surface is approximated through the node connectivity. If the position of a critical assembly point is adjusted after a design change, but the old version of the point is still being called, the system should prioritize reloading the assembly point coordinates according to the model version corresponding to the current production bill of materials. When the version cannot be automatically confirmed, the automatic scrap determination can be temporarily stopped, and only a manual re-inspection label can be output to prevent topology errors.

[0106] In the aforementioned inspection of the left front wheel arch, one sample showed a shallow anomaly on the outer side of the flange, while another sample showed a similar anomaly near the same arc surface of the upper buckle. If viewed only as a straight line in space, the distances of the two to the nearest buckle are similar. However, after calculating the shortest path along the curved surface, the geodesic distance from the latter to the root of the buckle along the continuous shell is smaller. Therefore, the system increases the distance attenuation weight of the latter and determines that it is more likely to affect the buckle pressing safety. In contrast, although the visual anomaly of the former is similar, the risk is appropriately reduced because it is located on the outer side of the flange and has a longer path connecting to the curved surface of the critical stress area.

[0107] The purpose of this mechanism is to assess the correlation between defects and critical assembly areas in a way that better reflects the actual stress path of the wheel arch structure, thereby achieving a physically reasonable weighting of the assembly failure risk index.

[0108] In this embodiment, the two-dimensional image feature extraction module includes a multi-protocol heterogeneous data acquisition adapter;

[0109] The multi-protocol heterogeneous data acquisition adapter reads multi-view two-dimensional surface image data through a preset industrial bus communication protocol;

[0110] The multi-protocol heterogeneous data acquisition adapter performs timestamp alignment and data format standardization processing on multi-view two-dimensional surface image data, and uniformly maps multi-view two-dimensional surface image data from different perspectives into multi-view two-dimensional surface image data in the global coordinate system of the inspection station.

[0111] This embodiment provides a heterogeneous image access and unified coordinate mapping mechanism suitable for mixed deployment environments of multiple devices in a workshop. Specifically, in the aforementioned online wheel cover inspection scenario, cameras and controllers from different production lines and different stages of modification may come from different manufacturers, and the output data protocols, time bases, and image formats are not consistent. If unified access and coordinate normalization are not completed first, although subsequent semantic segmentation can process individual images separately, it is difficult to guarantee that the multi-view images belong to the same data set at the same time, on the same workpiece, and under the same spatial reference, affecting the accurate mapping from two-dimensional coordinates to three-dimensional models. Therefore, this embodiment configures a multi-protocol heterogeneous data acquisition adapter in the two-dimensional image feature extraction module.

[0112] Specifically, the adapter reads surface image data from different camera nodes through a preset industrial bus communication protocol, and combines the clock signals from the station encoder, trigger sensor, or programmable logic controller to perform timestamp alignment on multi-view images. Its industrial significance lies in ensuring that the system processes a set of images of the same wheel cover at similar times but from different perspectives, rather than mixing images of two consecutive workpieces in the same group. The adapter performs format standardization processing on the image data, including resolution unification, color channel standardization, image naming rearrangement, and coordinate origin correction, so that images from different perspectives can be expressed in the same global coordinate system of the inspection station.

[0113] The so-called mapping refers to the multi-view image data under the global coordinate system of the inspection station. It can be understood as follows: Although each camera has its own local imaging coordinates, the system establishes the correspondence between the camera's local coordinates, the station's global coordinates, and the physical position of the workpiece through a calibration plate, workpiece reference holes, or fixture reference surfaces. Based on this, when a defect is identified as being located in a local area M1 of the image from a certain perspective, the system can know which segment of the workpiece's global coordinates corresponds to that area, avoiding conflicts between multi-view data. Subsequent two-dimensional coordinate extraction, three-dimensional mapping, and topological distance calculation are all based on this unified coordinate system.

[0114] In a simplified verification scenario, camera C1 can capture the outer surface of the wheel cover, and camera C2 can capture the inner surface of the wheel cover. The origin of the image output by C1 is located at the upper left corner, and the origin of the image output by C2 is located at the lower right corner, and the triggering time of the two differs by one sampling period. After processing by the adapter, the images of C1 and C2 are first matched to the same workpiece according to the beat signal, and then uniformly converted to the same naming rules and direction definitions. For example, the area originally marked as the upper left edge in C1 can fall into the same structural segment in the global coordinates of the workstation after conversion, along with the corresponding rib area in the view of C2, which is convenient for subsequent module fusion and judgment.

[0115] As an anomaly handling mechanism, if data from a certain perspective fails to arrive within a specified time window, the system can adopt different processing strategies based on the importance of the missing location. For non-critical observation surfaces, detection can continue while retaining the existing perspective, but the relevant areas will be marked as having incomplete perspectives. For critical areas such as the root of the buckle and the edge of the positioning pin hole, if the corresponding perspective is missing, the system will prioritize triggering re-acquisition. If re-acquisition fails, the workpiece will be directly transferred to the manual re-inspection channel. Furthermore, if camera calibration drift causes the unified coordinate mapping error to exceed the allowable range, the adapter can automatically block the input of that perspective based on the periodic calibration results to prevent erroneous coordinates from entering subsequent risk assessments.

[0116] During the upgrade and renovation of the same wheel cover inspection station, a new camera was added to acquire images of the inner side of the flange. Its communication protocol is different from that of the original outer surface camera, and the output image direction is also opposite. After the multi-protocol heterogeneous data acquisition adapter is connected to the camera, it first completes the reading according to the bus protocol, and then aligns the new camera image with the original multi-view image according to the same beat timestamp. After format standardization, it is incorporated into the global coordinate system of the station. As a result, the system can jointly analyze the newly discovered shrinkage area on the inner side of the flange with the dark lines on the outer surface under the original view, further improving the ability to identify the risk of hole edge warping.

[0117] The purpose of this mechanism is to solve the problem of image data being difficult to unify under multiple devices, protocols, and perspectives, thereby achieving reliable fusion of multi-perspective surface images within the same workstation coordinate framework, and providing a consistent data foundation for subsequent defect location and consequence classification.

[0118] 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 surface defect classification system for automotive wheel arch injection molded parts based on image recognition, characterized in that, include, The image acquisition module is used to acquire multi-view two-dimensional surface image data of the target component; The two-dimensional image feature extraction module performs semantic segmentation on the multi-view two-dimensional surface image data to generate an initial two-dimensional defect feature set containing surface defect morphological features and two-dimensional physical coordinates. The visual stress implicit mapping module calls a pre-built multimodal neural network model to map the initial two-dimensional defect feature set into the implicit residual stress estimate and local three-dimensional deformation probability of the target component. The multimodal neural network model is trained based on a pre-constructed joint training sample set, which includes injection mold flow analysis data based on computer-aided engineering and corresponding output physical object image data. The two-dimensional defect feature set extracted from the output physical object image data serves as the input features of the multimodal neural network model. The stress and deformation information of the regions in the injection mold flow analysis data corresponding to the output physical object image data serve as the supervision labels for the multimodal neural network model. The stress information includes the potential residual stress distribution, and the deformation information includes the warpage prediction region. The implicit residual stress estimate and the local three-dimensional deformation probability are the implicit physical inference results output by the multimodal neural network model after training with the joint training sample set. The implicit residual stress estimate corresponds to the potential residual stress distribution, and the local three-dimensional deformation probability corresponds to the warpage prediction region. The spatial topology tolerance calculation module acquires the 3D CAD model of the target component and the coordinate data of key assembly points. Based on the global coordinate system of the inspection station, it establishes a unified 2D and 3D spatial reference. Using the 2D physical coordinates, it determines the 3D mapping position of surface defects on the 3D CAD model and assigns a local 3D deformation probability. It calculates the topological distance from the 3D mapping position to the coordinate data of the key assembly points and constructs a distance attenuation weight that decreases as the distance increases. The estimated value of implicit residual stress is multiplied by this weight to obtain the basic stress risk value, which is then input into a preset exponential decay function and nonlinearly superimposed with the local 3D deformation probability. The preset exponential decay function includes a preset convergence coefficient to control the response convergence rate under high stress conditions. The nonlinear superposition amplifies the local 3D deformation probability based on a pre-calibrated warpage amplification factor, generating an assembly failure risk index. The dynamic consequence classification module outputs a fatal defect scrap label when the assembly failure risk index is greater than or equal to the preset safety tolerance upper limit; outputs a harmless defect release label when it is less than or equal to the preset safety tolerance lower limit; and outputs a manual re-inspection prompt label when it is between the lower and upper limits. The key assembly point coordinate data includes the three-dimensional coordinates of the buckle position and the three-dimensional coordinates of the locating pin position on the target component; The topological distance is specifically the shortest geodesic distance along the three-dimensional curved surface of the target component to the three-dimensional coordinates of the buckle position or the three-dimensional coordinates of the positioning pin position.

2. The image recognition-based surface defect classification system for automotive wheel arch injection molded parts according to claim 1, characterized in that, The surface defect morphological features include defect area data and texture gradient data; Before performing semantic segmentation, the two-dimensional image feature extraction module performs Gaussian filtering smoothing on the multi-view two-dimensional surface image data to remove high-frequency environmental noise signals from the multi-view two-dimensional surface image data.

3. The image recognition-based surface defect classification system for automotive wheel arch injection molded parts according to claim 1, characterized in that, The system also includes a process parameter fusion module; The process parameter fusion module acquires the core process parameters in the injection molding process; wherein, the core process parameters include mold temperature data and holding pressure time data; The visual stress implicit mapping module performs a tensor concatenation operation on the core process parameters and the initial two-dimensional defect feature set, which serves as the joint input data for the multimodal neural network model.

4. The image recognition-based surface defect classification system for automotive wheel arch injection molded parts according to claim 1, characterized in that, The system also includes a feedback correction module; The feedback correction module acquires actual assembly interference feedback data from the subsequent assembly stage. The feedback correction module compares the actual assembly interference feedback data with the fatal defect scrap label or harmless defect release label output by the dynamic consequence classification module. When the comparison result is inconsistent, the feedback correction module generates a model correction instruction to drive the visual stress implicit mapping module to backpropagate and update the network weights of the multimodal neural network model. When the comparison result is consistent, the feedback correction module generates a state maintenance instruction to maintain the current network weights of the multimodal neural network model.

5. The image recognition-based surface defect classification system for automotive wheel arch injection molded parts according to claim 1, characterized in that, The two-dimensional image feature extraction module includes a multi-protocol heterogeneous data acquisition adapter; The multi-protocol heterogeneous data acquisition adapter reads the multi-view two-dimensional surface image data through a preset industrial bus communication protocol; The multi-protocol heterogeneous data acquisition adapter performs timestamp alignment and data format standardization processing on the multi-view two-dimensional surface image data, and uniformly maps the multi-view two-dimensional surface image data from different perspectives into multi-view two-dimensional surface image data in the global coordinate system of the detection station.

Images