Visual analysis-based inspection device for automotive stamping parts

Abstract

This invention discloses a vision-based inspection device for automotive stamped parts, relating to the fields of non-destructive testing and opto-mechatronics (OMT) technology. The device includes a frame, a workpiece conveying device, a programmable light source device, a vision inspection device, a defective part rejection device, and a control cabinet. It acquires surface images of stamped parts using an industrial camera, divides them into grid blocks, and extracts reflection interference and morphological anomaly indices. It combines servo fan load current signals to obtain vacuum leakage mutation factors and retrieves offline forming limit diagrams to obtain critical cracking risk coefficients, constructing an FLD modulation coupled damage index model for defect judgment. Based on the judgment results, the system performs closed-loop light field scheduling and in-situ re-image verification, and coordinates with the defective part rejection device to complete sorting. This device can effectively suppress interference such as oil film reflection, improve defect identification accuracy, and enhance inspection stability.

Description

Technical Field

[0001] This invention relates to the fields of nondestructive testing and opto-mechatronics technology, and more specifically, to a visual analysis-based inspection device for automotive stamping parts. Background Technology

[0002] Currently, in the field of visual analysis-based inspection of automotive stamped parts, although existing technologies have explored image algorithms, 3D imaging, and mechanical multi-angle scanning, it can be seen from the three comparative documents CN114627080B, CN117291918B, and CN108519389A that these solutions still have significant shortcomings and are difficult to fully support the demanding industrial-grade online full inspection business with high cycle times, complex curved surfaces, and strong reflective interference.

[0003] Firstly, the vehicle stamping parts defect detection solution represented by CN114627080B focuses on extracting the grayscale values ​​of each pixel in the image. It judges surface defects by calculating the degree of grayscale fluctuation between adjacent pixels and the wrinkle heat of the fitted surface. This type of method solves more about the feature extraction and post-processing of static images, rather than the fundamental control of the physical imaging environment and light field interference. Its feature extraction is highly dependent on the static two-dimensional image acquired in a single shot, lacks the means to remove high reflective interference, and does not introduce a coupling judgment mechanism between morphological anomalies and reflection intensity. This leads to the image pixels appearing white due to irreversible loss of physical information when the complex curved surface or oil film of the stamping part causes high light overexposure. The grayscale fluctuation extraction at the pure algorithm level completely fails, which easily produces false positives or causes the real defects to be covered by reflection. These are precisely the problems that are unacceptable in the quality inspection process of automotive parts manufacturing, where there is zero tolerance for missed detection rates. Secondly, CN117291918B proposes a defect detection method for automotive stamping parts based on 3D point clouds. Its main advantage lies in the detection of latent defects such as slight thickness variations. It uses point cloud data acquired by a 3D camera to overcome the visual blind spots of a 2D plane. However, the positioning of this solution is still the up-dimensional acquisition of data. The generation of its point cloud is based on the assumption that the optical surface exhibits ideal diffuse reflection. Its hardware platform relies heavily on high-precision static scanning. It does not have fine-grained anti-interference methods for optical interferences such as high-frequency refraction of the drawing oil film and specular reflection on the surface of stamping parts. It also lacks a two-way correction mechanism for predicting the physical normal changes caused by reflection. When facing complex curved metal parts coated with lubricant in a real workshop, strong reflection will cause serious voids or distortions in the point cloud calculated by the 3D camera. Furthermore, the solution does not provide a design for the underlying light field reconstruction to dynamically eliminate such physical interferences, making it difficult to directly apply the solution to real-time flow shooting lines that require millisecond-level response and are subject to severe oil contamination interference. CN108519389A uses a complex mechanical slide, slider, and electric telescopic rod to adjust the spatial displacement of the CCD camera and the light source. While this design improves the coverage of the multi-angle field of view and reduces fixed blind spots to some extent, its mechanical adjustment is more used for coarse positioning and multi-station workflow before inspection. It does not refine the control of the light source to a granular level that provides real-time feedback mapping with the reflectivity or morphological features of the image surface, nor does it specify a targeted light field reconstruction mechanism that is automatically triggered based on image quality. When dealing with workpieces with extremely complex structures and rapidly changing reflective surfaces, mechanical actions may still cause problems. The lag can lead to risks such as lighting mismatch and local overexposure. Furthermore, the three solutions generally lack a system design that addresses the physical coupling and closed-loop feedback of both reflection interference and morphological anomalies. Specifically, when severe optical interference is detected in a region, how to quantify the degree to which the interference masks morphological features, how to use algorithms to reverse-guide the illumination angle and flicker of the programmable light source, and under what conditions to generate interference-free target images are all unresolved systematically. This results in the system being forced to passively report errors or rely on manual re-inspection in complex detection environments, significantly increasing the risk of missed detections and the maintenance costs of the production line.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a visual analysis-based automotive stamping parts inspection device. By introducing flexible negative pressure adsorption to establish a stable physical benchmark, and constructing an FLD modulation coupled damage index model that integrates visual optical features, aerodynamic leakage factors, and CAE forming limits, combined with the dynamic closed-loop scheduling of a programmable light source, the device can accurately distinguish between real defects and optical artifacts on highly reflective surfaces, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A vision-based automotive stamping parts inspection device includes a frame, a workpiece conveying device, a programmable light source device, a vision inspection device, a defective product rejection device, and a control cabinet. The workpiece conveying device is horizontally mounted in the lower middle part of the frame, extending through the inspection area, and includes a servo-frequency inverter fan. The programmable light source device is suspended directly above the inspection area. The vision inspection device adopts a zenith-coaxial vertical top-down layout and includes an industrial camera. The defective product rejection device is located at the discharge end of the workpiece conveying device along the workpiece conveying direction. The control cabinet contains an edge computing industrial control computer. Based on the above structure, the edge computing industrial control computer receives image data acquired by the industrial camera and divides it into local grid blocks to obtain the reflection interference index and shape anomaly index of each block. The edge computing industrial control computer is also used to determine the vacuum leakage mutation factor based on the load current signal of the servo variable frequency fan drive circuit, and extract the critical cracking risk coefficient of the corresponding block based on the offline CAE forming limit diagram; the edge computing industrial control computer constructs the FLD modulation coupling damage index model, and inputs the reflection interference index, the morphological anomaly index, the vacuum leakage mutation factor and the critical cracking risk coefficient as detection parameters into the FLD modulation coupling damage index model to obtain the defect judgment index; the edge computing industrial control computer will perform dynamic light field scheduling control on the programmable light source device to realize light field reconstruction based on the spatial distribution of the defect judgment index, and perform in-situ re-shooting through the industrial camera to obtain the reconstructed light field image for secondary analysis, and link the defect rejection device for physical sorting when a defect is found.

[0007] In a preferred embodiment, the frame includes: a base beam, damping and shock-absorbing feet, a gantry column, and a top support platform; multiple base beams are orthogonally welded to each other to form a rectangular chassis structure, and the damping and shock-absorbing feet are screwed to the four corners of the bottom; the gantry column extends vertically upward from the end of the base beam and is rigidly connected to the top support platform at the top.

[0008] In a preferred embodiment, the workpiece conveying device further includes: a flexible negative pressure conveyor belt, a servo drive motor, a driving roller, a driven tensioning roller, a vacuum adsorption box, an absolute encoder, and a laser beam sensor; the axes of the driving roller and the driven tensioning roller are both parallel to the horizontal plane and perpendicular to the workpiece conveying direction; the servo drive motor is coaxially connected to one end of the driving roller; the absolute encoder is coaxially connected to the other end of the driving roller; and the laser beam sensor spans both sides of the conveying path at the feeding end.

[0009] In a preferred embodiment, the vacuum adsorption box is embedded between the active roller and the driven tensioning roller, with its upper end face open and closely attached to the bottom surface of the flexible negative pressure conveyor belt, and its lower end sealed and connected to the servo variable frequency fan at the bottom through a flexible air pipe; the surface of the flexible negative pressure conveyor belt is provided with a soft coating with a high coefficient of friction, and the surface of the flexible negative pressure conveyor belt is uniformly distributed with an array of micropores.

[0010] In a preferred embodiment, the upper surface of the vacuum adsorption box is provided with a low-friction support plate with a flow-guiding grid groove; a gas-liquid separation filter is connected in series between the flexible air pipe at the bottom of the vacuum adsorption box and the servo frequency converter fan; a current sampling module is provided in the drive circuit of the servo frequency converter fan to collect the fan load current signal in real time and upload it to the edge computing industrial control computer.

[0011] In a preferred embodiment, the programmable light source device specifically includes: a hemispherical light source cover, a visible light LED array, a polarized light LED array, an ultraviolet LED array, a heat dissipation conformal plate, an adjustment frame, adjustment bolts, and a light source control circuit board; the visible light LED array, the polarized light LED array, and the ultraviolet LED array are arranged in concentric circles, alternately nested, on the concave hemispherical surface of the hemispherical light source cover; the heat dissipation conformal plate is tightly fitted to the convex spherical surface of the hemispherical light source cover.

[0012] In a preferred embodiment, one end of the adjustment frame is fixed to the top support platform, and the other end is connected to the hemispherical light source cover. The adjustment frame is provided with the adjustment bolt, so that the hemispherical light source cover can be finely adjusted in the vertical direction.

[0013] In a preferred embodiment, the geometric apex of the hemispherical light source cover has a light-transmitting hole, and the visual inspection device further includes: a large depth-of-field dual telecentric lens, a vertical high-precision fine-tuning slide, and a polarizing filter; the vertical high-precision fine-tuning slide is vertically mounted on the outside above the light-transmitting hole; after the industrial camera and the large depth-of-field dual telecentric lens are assembled, they are mounted on the vertical high-precision fine-tuning slide, and the polarizing filter is screwed to the front end of the lens; the main optical axis of the industrial camera is strictly parallel to the vertical direction and passes through the center of the hemisphere.

[0014] In a preferred embodiment, the control cabinet is independently mounted to isolate the production line from vibration, and its interior also includes a high-voltage distribution module; the defective product rejection device is located on the side of the discharge end, and includes a high-speed solenoid valve and a pneumatic push rod, the control end of which is electrically connected to the edge computing industrial control computer, and is used to push out defective stamped parts laterally.

[0015] In a preferred embodiment, the control cabinet integrates an edge control hub, which includes: an image acquisition unit for receiving an initial light field image of the stamped part surface acquired by an industrial camera, and a reconstructed light field image captured in situ after closed-loop light field adjustment; a feature analysis unit for dividing the acquired initial or reconstructed light field image into local grid blocks, calculating the optical and geometric properties of each block in parallel, and outputting a reflection interference index and a shape anomaly index; and a coupling decision unit for constructing an FLD modulation coupling damage index model, retrieving an offline CAE forming limit map to extract the critical cracking risk coefficient of the corresponding block, and receiving data from the edge computing industrial control computer based on the servo variable frequency fan load current. The vacuum leakage mutation factor obtained from signal calculation is used to jointly couple the reflection interference index, the morphological anomaly index, the critical cracking risk coefficient, and the vacuum leakage mutation factor to obtain a defect judgment index, thereby generating a global defect probability distribution field. The light source control unit, with an embedded topology mapping table, is used to receive the global defect probability distribution field signal output by the coupled decision unit, and when it is determined that light field optimization is required, it starts the dynamic light field adaptive control model and outputs spatial coordinate control commands to the light source control circuit board to achieve light field reconstruction by adjusting the spatial illumination distribution of the programmable light source device in a closed loop. The alarm output unit is used to output high-confidence defect coordinates and defect types after the light field closed-loop adjustment is completed.

[0016] The technical effects and advantages of the vision analysis-based automotive stamping parts inspection device of this invention are as follows: This invention effectively suppresses false artifacts by constructing an adaptive optical field closed-loop scheduling mechanism. Combined with a multispectral programmable light source mechanism and a high-precision fine-tuning slide, it automatically triggers an optical field reconstruction strategy for suspected defect areas, accurately stripping away non-structural optical interference such as drawing oil film and surface particles, significantly reducing the probability of quality inspection misjudgment and reducing production line intervention. This invention establishes an FLD modulation coupled damage index model to achieve high sensitivity in distinguishing potential defects. It effectively combines visual features with the risk coefficient of offline forming limit map, uses the mechanical limit of materials as a weight constraint, automatically suppresses light field disturbance in the safe forming area, and nonlinearly amplifies the defect response in the high-risk cracking area, effectively enhancing the system's ability to distinguish real defects. This invention improves detection reliability by incorporating a pneumatic and visual cross-physical field collaborative mechanism. The flexible negative pressure conveying mechanism provides high-rigidity imaging support for thin-walled parts, reduces deformation interference, and extracts the vacuum leakage mutation factor of the servo fan load current in real time and incorporates it into the model. When penetrating damage occurs, the underlying aerodynamic anomaly is directly used as a redundant criterion to enhance the alarm, effectively avoiding malicious missed detections. Attached Figure Description

[0017] Figure 1This is a schematic diagram of the overall structure of the automotive stamping parts inspection device based on visual analysis according to the present invention.

[0018] Figure 2 This is a schematic diagram of the workpiece conveying device of the present invention.

[0019] Figure 3 This is a schematic diagram of the structure of the programmable light source device and the visual inspection device of the present invention.

[0020] Figure 4 This is a schematic diagram of the multispectral LED array distribution inside the hemispherical light source cover of the present invention.

[0021] Figure 5 This is a block diagram of the edge control hub system architecture of the present invention.

[0022] Figure 6 This is a flowchart of the detection method based on the FLD modulation coupling damage index model of the present invention.

[0023] Figure 7 This is a logic diagram for judging the optical field optimization state based on the defect judgment index in this invention.

[0024] Figure 8 This is a comparison chart of the defect judgment index response under typical working conditions of the present invention.

[0025] 1. Frame; 2. Base beam; 3. Damping and vibration-damping feet; 4. Gantry column; 5. Top support platform; 6. Stamping parts; 7. Flexible negative pressure conveyor belt; 8. Servo drive motor; 9. Drive roller; 10. Driven tension roller; 11. Vacuum adsorption box; 12. Servo frequency converter fan; 13. Gas-liquid separation filter; 14. Absolute encoder; 15. Laser beam sensor; 16. Pneumatic push rod; 17. Waste collection trough; 18. Hemispherical light source cover; 19. Visible light LED array; 20. Polarized light LED array; 21. Ultraviolet LED array; 22. Heat dissipation conformal plate; 23. Adjustment frame; 24. Adjustment bolt; 25. Light source control circuit board; 26. Industrial camera; 27. Large depth-of-field double telecentric lens; 28. Polarizing filter; 29. ​​Vertical high-precision fine-tuning slide; 30. Control cabinet. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Example 1, as Figures 1 to 4As shown, this embodiment discloses a vision analysis-based inspection device for automotive stamped parts. This device combines machine vision with opto-mechatronics collaborative control to achieve automatic inspection of the surface quality of stamped parts. For ease of spatial description, this embodiment defines the conveying direction of the stamped parts as the positive X-axis, the horizontal plane perpendicular to the conveying direction as the Y-axis, and the vertically upward direction as the Z-axis.

[0028] The device adopts a high-rigidity gantry anti-vibration structure, which includes, from bottom to top, a frame 1, a workpiece conveying device, a programmable light source device, a vision inspection device, a defective product rejection device, and an external control cabinet 30. The rack 1 serves as the physical reference carrier for the entire opto-mechatronics system, specifically including: base beam 2, damping and shock-absorbing feet 3, gantry column 4, and top support platform 5; Furthermore, multiple base beams 2 are orthogonally welded together to form a rectangular chassis structure, with damping and vibration-damping feet 3 screwed to the four corners of the bottom. The gantry column 4 extends vertically along the Z-axis from the end of the base beam 2 and is rigidly connected at the top via a top support platform 5; the damping and vibration-damping feet 3 have built-in heavy-duty rubber springs, and by adjusting the feet, the high-precision level of the top plane of the frame 1 is ensured, effectively isolating the low-frequency resonance transmitted from the workshop floor by the large punch press, thereby improving the stability of the optical imaging reference plane.

[0029] The workpiece conveying device is horizontally mounted in the lower middle part of the frame 1, running through the detection area. Specifically, it includes: a flexible negative pressure conveyor belt 7, a servo drive motor 8, an active roller 9, a driven tension roller 10, a vacuum adsorption box 11, a servo frequency conversion fan 12, an absolute encoder 14, and a laser beam sensor 15. Preferably, the surface of the flexible negative pressure conveyor belt 7 is provided with a soft coating with a high coefficient of friction, such as a soft polyurethane or silicone coating. Compared with ordinary PVC belts, this soft coating material can produce slight elastic deformation under negative pressure suction, thereby adapting to and conforming to the local irregular edges of the stamped parts to be inspected to the maximum extent, effectively increasing sealing resistance, preventing negative pressure leakage, and ensuring stable adsorption and accurate positioning of the stamped parts during visual inspection. Furthermore, the axes of both the driving roller 9 and the driven tensioning roller 10 are parallel to the Y-axis. The servo drive motor 8 is coaxially connected to one end of the driving roller 9, and the absolute encoder 14 is coaxially connected to the other end of the driving roller 9. The vacuum adsorption box 11 is embedded between the driving and driven rollers, with its upper end open and tightly attached to the bottom surface of the flexible negative pressure conveyor belt 7. Its lower end is sealed and connected to the servo frequency converter fan 12 at the bottom through a flexible air pipe. The laser beam sensor 15 spans both sides of the X-axis feed end. When the laser beam sensor 15 captures the leading edge of the workpiece, it triggers the servo frequency converter fan 12 to establish a transient negative pressure in the vacuum adsorption box 11, firmly adsorbing the thin-walled stamped part onto the surface of the flexible negative pressure conveyor belt 7, suppressing the slight warping and vibration of the stamped part during the conveying process. The absolute encoder 14 picks up the mechanical angular displacement of the driving roller 9 in near real-time and outputs a pulse signal corresponding to the physical displacement of the X-axis, providing a spatial coordinate reference for the entire system. Furthermore, the upper surface of the vacuum adsorption box 11 is provided with a low-friction support plate with a guide grid groove, and the surface of the flexible negative pressure conveyor belt 7 is evenly distributed with arrayed micropores; a gas-liquid separation filter 13 is connected in series between the flexible air pipe at the bottom of the vacuum adsorption box 11 and the servo frequency conversion fan 12. Since the surface of the stamped part to be tested is usually covered with drawing oil film and workshop dust, the gas-liquid separation filter 13 is used to intercept and settle liquid oil droplets and solid particles in the airflow before the negative pressure suction airflow enters the fan, thereby protecting the impeller and motor of the servo frequency conversion fan 12 from contamination and damage, and ensuring the long-term stable operation of the negative pressure system. The servo frequency conversion fan 12 is configured as a high-flow centrifugal fan to overcome the normal air leakage at the edge of the three-dimensional curved surface of the stamped part through dynamic flow compensation, ensuring the dynamic adsorption rigidity of the detection plane; in one embodiment, a current sampling module is provided in the drive circuit of the servo frequency conversion fan 12 to collect the fan load current signal in real time and upload it to the edge computing industrial control computer for calculating the vacuum leakage mutation factor.

[0030] The programmable light source device is suspended directly above the detection area and specifically includes: a hemispherical light source cover 18, a visible light LED array 19, a polarized light LED array 20, an ultraviolet LED array 21, a heat dissipation conformal plate 22, an adjustment frame 23, an adjustment bolt 24, and a light source control circuit board 25; the ultraviolet LED array 21 preferably uses a wavelength of 365nm to excite the fluorescence characteristics of the drawn oil film on the surface of the stamped part; Furthermore, the upper end of the adjustment frame 23 is fixed to the top support platform 5, and the lower end is connected to the hemispherical light source cover 18. The adjustment frame 23 is equipped with adjustment bolts 24, which allow the hemispherical light source cover 18 to be finely adjusted in the Z-axis direction. The visible light LED array 19, the polarized light LED array 20, and the ultraviolet LED array 21 are arranged in concentric circles on the concave hemispherical surface of the light source cover. The heat dissipation conformal plate 22 is tightly fitted to the convex spherical surface of the hemispherical light source cover 18. The adjustment frame 23 ensures that the lower edge of the light source cover maintains a precise safety gap with the conveyor belt. The alternating arrangement of the visible light LED array 19, the polarized light LED array 20, and the ultraviolet LED array 21 can highlight the differences in surface texture of the stamped parts under different irradiation conditions, thereby enhancing the imaging contrast of fine scratches and dents. The heat dissipation conformal plate 22 reduces the risk of spectral thermal drift.

[0031] The visual inspection device adopts a zenith coaxial vertical top-down layout, specifically including: an industrial camera 26, a large depth-of-field dual telecentric lens 27, a vertical high-precision fine-tuning slide 29, and a polarizing filter 28. Furthermore, a light-transmitting hole is provided at the geometric apex of the hemispherical light source cover 18, and a vertical high-precision fine-tuning slide 29 is vertically mounted on the outside above the light-transmitting hole. After the industrial camera 26 and the large depth-of-field dual telecentric lens 27 are assembled, they are mounted on the vertical high-precision fine-tuning slide 29, and a polarizing filter 28 is screwed onto the front end of the lens. The main optical axis of the industrial camera 26 is strictly parallel to the Z-axis and passes through the center of the hemisphere; the vertical high-precision fine-tuning slide 29 is used to precisely calibrate the focal plane of the industrial camera 26. The polarizing filter 28 and the polarized light LED array 20 below form an orthogonal optical path coupling. The bottom pins of the industrial camera 26 are directly connected to the absolute encoder 14 and the light source control circuit board 25 to realize hardware-level synchronous triggering between the light source flicker and the exposure of the industrial camera 26; Furthermore, to ensure that this device meets the millisecond-level rapid-fire full-inspection requirements for highly reflective stamped parts, this embodiment explicitly defines the core hardware specifications: the industrial camera 26 is preferably a global shutter CMOS large-area array camera with a resolution of not less than 65 million pixels and a full frame rate. To eliminate motion blur; the matching large depth-of-field dual telecentric lens has a preferred focal length of 110mm-130mm, providing a depth of field. Field of view coverage This ensures that all the undulating surfaces of the stamped parts are in focus on a clear focal plane. The polarization direction of the polarizing filter 28 is strictly orthogonal to the preset polarization direction of the polarizing LED array 20. The underlying servo variable frequency fan 12 is preferably a high-flow-rate, high-pressure centrifugal fan with a maximum air extraction flow rate. Ultimate vacuum This ensures that the flexible negative pressure conveyor belt maintains a stable absolute physical reference at a linear speed of 60m / min.

[0032] The defective product rejection device is located on the side of the discharge end of the workpiece conveying device along the positive X-axis. The defective product rejection device includes a high-speed solenoid valve and a pneumatic push rod 16, and its control end is electrically connected to the edge computing industrial control computer. When the edge computing industrial control computer determines that the current stamped part has a surface defect, it performs delayed tracking based on the displacement data recorded by the absolute encoder 14. When the defective workpiece reaches the rejection station, the high-speed solenoid valve is triggered, and the pneumatic push rod 16 pushes it laterally into the waste collection tank 17, realizing automatic physical sorting of good and defective products.

[0033] The control cabinet 30 is installed separately to isolate the vibration of the production line. The control cabinet 30 includes an edge computing industrial control computer and a high power distribution module. The edge computing industrial control computer can receive image data collected by the industrial camera 26, execute visual analysis algorithms, and send control commands to the light source control circuit board 25 and the servo drive motor 8 according to the analysis results, and perform image analysis and light field adjustment during the detection process.

[0034] Based on the above structure, the workpiece enters the detection area along the X-axis under the drive of the workpiece conveying device. The absolute encoder 14 outputs the conveying displacement information in real time as the system spatial coordinate reference. The programmable light source device forms a multi-angle light field covering the detection area under the action of the control signal. The vision inspection device acquires the surface image of the stamped part under hard synchronization triggering conditions. The edge computing industrial control computer in the control cabinet 30 performs real-time analysis on the acquired image and controls the light source array and conveying status according to the analysis results. When a defect is found, it links the defective product rejection device to perform physical sorting, thereby forming a closed loop and finally completing the location identification, type determination and automatic rejection of defective products on the surface of the stamped part.

[0035] Example 2, as Figure 5 As shown, the control cabinet 30 integrates an edge control hub, which constructs an intelligent linkage mechanism for feature tensor extraction, FLD modulation coupling impairment index model, and optical field closed-loop reconstruction. The edge control hub includes: An image acquisition unit is used to receive the initial light field image and reconstructed light field image of the stamping surface acquired by the industrial camera 26 through a hard synchronization interface. The feature analysis unit, which embeds an FPGA-accelerated tensor extractor, is used to divide the acquired image into local grid blocks. The optical and geometric properties of each block are calculated in parallel, and the reflection interference index is output. and morphological abnormality index ; Furthermore, the edge computing industrial control computer adopts a heterogeneous computing architecture of a central processing unit (CPU), a graphics processing unit (GPU), and a field-programmable gate array (FPGA). Specifically, the industrial control computer is preferably configured with a 16-core industrial-grade CPU with a base frequency of not less than 3.0 GHz, a GPU with not less than 4,000 stream processor cores and 16 GB of video memory, and not less than 64 GB of error-correcting memory. The FPGA-accelerated tensor extractor preferably uses an industrial-grade FPGA with not less than 500,000 logic units and not less than 2,500 digital signal processing multiply-accumulator slices. The system is a heterogeneous system-on-a-chip (SoC). Data collected by the industrial camera 26 is written at high speed to the on-chip block memory of the FPGA via a gigabit industrial vision protocol and a direct memory access mechanism. The FPGA internally instantiates a pipelined digital signal processing multiplier-accumulator array, which performs second-order Hessian matrix differentiation and Sobel gradient magnitude convolution operations in full parallel at the low-level hardware level of the clock cycle. After completing the basic pixel-level calculation, the condensed feature tensor is pushed to the main computing unit of the edge computing industrial control computer via a high-speed bus for floating-point operations of the FLD modulation coupled damage index model.

[0036] The coupling decision unit is used to construct the FLD modulation coupling impairment index model and to convert the reflection interference index into a coupling decision unit. With morphological abnormality index As a basic characteristic, and combined with the FLD risk coefficient With vacuum leakage mutation factor Perform coupled calculations; The light source control unit has an embedded topology mapping table, which is used to receive the global defect probability distribution field signal output by the coupled decision unit, and to start the dynamic light field adaptive control model when it is determined that light field optimization is required, and output spatial coordinate control commands to the light source control circuit board 25 to adjust the spatial illumination distribution of the LED light source array in a closed loop. An alarm output unit is used to output the defect coordinates and defect type with high confidence after the optical field closed-loop adjustment is completed.

[0037] Example 3: This example proposes a vision analysis-based inspection method for automotive stamped parts, relying on the aforementioned detection device and control system. Figure 6 As shown, by constructing an FLD modulation coupled damage index model, effective distinction can be made between real defects and optical artifacts. Furthermore, by combining this with a programmable light source device to achieve dynamic closed-loop adjustment of the optical field, the detection stability and accuracy are significantly improved. The method specifically includes the following steps: Step S1: When the stamped part enters the visual inspection area along the X-axis direction with the conveyor belt, the system's detection input is not initiated by pure software recognition, but is directly assigned an absolute spatiotemporal reference by physical hardware. Specifically: when the physical beam of the laser beam sensor 15 spanning both sides of the conveyor belt is cut off by the leading edge of the workpiece, the system's detection time reference is triggered; simultaneously, the absolute encoder 14, coaxially connected to the non-drive end of the drive roller 9, synchronously outputs physical displacement pulses as subsequent image grid blocks. The absolute spatial addressing reference; At the moment the detection signal is triggered, the vacuum adsorption box 11 embedded inside the conveyor loop establishes a transient negative pressure, which rigidly adsorbs the thin-walled stamped part onto the surface of the conveyor belt. In conjunction with the damping and shock-absorbing feet 3 of the base, it isolates the low-frequency resonance of the workshop, thereby providing a stable physical imaging reference for subsequent steps. Then, the programmable light source device turns on the reference illumination field, and the industrial camera 26 completes global image acquisition.

[0038] In step S2, the feature analysis unit divides the acquired global image into regular grids, forming multiple local analysis blocks. For each block, the system strictly extracts two types of basic visual feature parameters, and the depth of feature extraction depends on the customized hardware architecture of this device. Furthermore, the global image is mapped to local grid blocks. The partitioning is not based on a fixed size, but on a curvature-guided adaptive partitioning rule: after the edge computing industrial control computer acquires the image, it retrieves the surface curvature gradient matrix of the stamping part's CAD model; in flat areas, the block size is kept at the maximum first-level size to accelerate the calculation; in high curvature feature areas with abrupt changes in R-corners, folded edges, and drawing depth, the system triggers a quadtree splitting algorithm to split down to the minimum physical resolution limit; The block merging rule is as follows: if the angle between the gray-level variance and the shape gradient vector of four adjacent sub-blocks of the same level is less than 5%, then reverse merging is triggered. This dynamic partitioning mechanism takes into account both the extremely high spatial resolution of feature extraction and the computational efficiency of whole-image solution.

[0039] Reflection Interference Index Used to characterize the distribution and high reflectivity of the drawn oil film on the surface of stamped parts, and to quantify the oil film accumulation artifacts; the multimodal light field input directly depends on the visible light and ultraviolet LED array 21 arranged in concentric nested circles on the inner wall of the hemispherical light source cover 18; under hard synchronization triggering, the system extracts the oil film fluorescence reflectance excited in a specific ultraviolet band, and calculates the ratio between it and the visible light reflectance, thereby accurately quantifying the intensity of pure optical artifacts caused by the drawn oil accumulation; Morphological Anomaly Index Used to characterize geometric abrupt changes on the surface of stamped parts, ridge features and gradient amplitudes are extracted. Thanks to the absolutely high rigidity physical benchmark provided by the vacuum adsorption system and the shock-resistant frame 1 in step S1, the system can stably extract local curvature change features using the ridge feature analysis method based on the second derivative matrix, and combine it with the edge gradient amplitude to obtain an extremely sensitive and physical artifact-free morphological anomaly index. .

[0040] Step S3, the coupled decision unit sets the reflection interference index. With morphological abnormality index Joint coupling calculations are performed, and the system synchronously retrieves the offline CAE forming limit diagram of the batch of stamped parts, and extracts the critical cracking risk coefficient of the corresponding block. A FLD modulation coupled damage index model was constructed to evaluate the coupling effect of structural roughness evolution, deformation characteristics, and forming limit safety margin on damage evolution, thereby obtaining a defect judgment index that integrates the genes of stamping process. The preferred calculation formula is: ; in: Indicates the vacuum leakage mutation factor; This represents the critical cracking risk coefficient of the FLD (Fluorescent Delay) in the preloaded automotive stamping forming limit diagram in this block, i.e., the FLD risk coefficient. The reflection basis weight coefficient is preferably in the range of 0.1 to 0.3, which characterizes the system's tolerance to purely optical interference. Its value is inversely proportional to the basic reflectivity of the material of the current batch of stamped parts. The morphology-based weighting coefficient, preferably ranging from 0.4 to 0.7, characterizes the system's sensitivity to physical geometric distortions and is used to maintain the system's basic detection capability for subtle deformations. The nonlinear coupling gain coefficient is preferably in the range of 1.5 to 4.0. It is the core control parameter of the energy function and is used to control the exponential amplification when reflection and morphology undergo abrupt changes. In the FLD modulation coupling damage index model, the first term characterizes the intensity of optical reflection interference in the local region, the second term characterizes the degree of abnormality in the surface morphology and structure, and the third term is the coupling enhancement term.

[0041] Furthermore, vacuum leakage mutation factor This is a unique parameter of the opto-mechatronics collaborative architecture of the present invention; during the detection process, the edge computing industrial control computer monitors the underlying torque load current of the servo variable frequency fan 12 in near real-time; when the stamped part has surface scratches that are not completely torn, the vacuum adsorption box 11 maintains good pressure. The value is set to a baseline of 1.0; when a localized through-crack causes transient micro-leakage, the servo load slightly decreases. The value will increase adaptively and dynamically; the introduction of this aerodynamic hardware parameter enables the coupling enhancement term to accurately distinguish between surface defects and through tears, achieving cross-physical field verification. Furthermore, to ensure the vacuum leakage mutation factor With accurate extraction, this invention clarifies the underlying current sampling standard and adaptive quantization formula for servo variable frequency fans; The current sampling module employs a high-speed ADC with a resolution of at least 16 bits to perform high-frequency sampling of the actual torque current on the DC bus side of the servo variable frequency fan driver. The sampling frequency is preferably set to 10kHz. To eliminate interference from high-frequency harmonics in the workshop power grid and PWM switching noise from the inverter, the edge computing industrial control computer, after receiving the raw current signal, first performs preprocessing using moving average filtering and Kalman filtering algorithms to extract a smooth, real-time effective value of the load current. ; The edge computing industrial control computer performs real-time calculations based on the following adaptive mathematical model. : ; in: This indicates the reference load current value extracted by the system through the calibration program under the adsorption and pressure holding state of the normal good stamped parts in this batch; This represents the preset dynamic dead zone threshold, preferably ranging from 0.02 to 0.05. This threshold is set adaptively based on the natural leakage rate at the edge of the batch of stamped parts and the normal fluctuation range of the power grid. Dynamic Dead Zone Threshold It filters out current disturbances caused by non-defect factors, ensuring It remains absolutely stable at 1.0 under normal operating conditions; This represents the leakage gain coefficient, with a preferred value range of 5.0 to 10.0; The nonlinear sensitivity index is preferably fixed at 1.5. Based on the above quantitative rules, when a non-penetrating defect or normal fluctuation occurs on the surface of the stamped part, the rate of change of current is insufficient to exceed the dead zone threshold. , The baseline value of 1.0 is maintained, without interfering with visual coupling determination; however, once a penetrating tear occurs, the current drop rate exceeds [a certain threshold]. The slight leakage signal will be rapidly amplified by the nonlinear amplification term. For example, when the current drops slightly by 8% and is set... , , When, the calculation yields At this point, the aerodynamic mutation factor is directly injected as a second-order gain multiplier into the FLD modulation coupled damage index model, realizing a strong endorsement of the visual algorithm at the physical level, thereby completely opening up the collaborative judgment mechanism across physical fields.

[0042] Furthermore, the Forming Limit Diagram (FLD) is a core metallurgical mechanical tool used in the automotive stamping field to evaluate the sheet metal drawing performance. In this embodiment, the system pre-imported CAE forming simulation data of this batch of stamped parts during the early mold development stage and aligned it with the spatial coordinate system of visual inspection in three dimensions. The specific numerical value represents the current visual grid block. In physics and mechanics, the greater the value of the distance from the critical point of necking or cracking, the higher the a priori mechanical probability of the presence of microcracks in that region. Preferably, in order to make the defect determination index The dimensionless measurement of the FLD risk coefficient Before participating in the coupled calculation, the principal and secondary strain data output by the edge computing industrial control computer based on CAE simulation are normalized and mapped. The value range is preferably limited to [0, 1.5]. Furthermore, according to the physical deformation state of automotive stamping parts, the system strictly divides this coefficient into three stepped response intervals: when The safe forming range is defined as the region where the actual strain of the grid block is much lower than the forming limit curve, indicating that the region does not meet the physical and mechanical conditions for drawing cracking. At this point, the smaller... The numerical value will suppress the coupling enhancement term. Even if there is a large area of ​​stretching oil film in the region causing reflection interference, the system will not trigger a defect alarm, thus effectively filtering out optical artifacts caused by stamping oil film at the algorithm level. when The time interval is defined as the necking warning zone. Within this zone, the strain of the grid block gradually approaches the forming limit curve, indicating that significant lattice slip has occurred within the metal material in this region, posing a potential risk of necking. The linear weighting coefficients, which serve as coupling enhancement terms, begin to play a role, making the system more sensitive to the optical and morphological distortion characteristics of the region, thereby enabling early warning of potential microcracks or local tearing risks. when Time is defined as the critical cracking and rupture interval, when When the value approaches or exceeds 1.0, it indicates that the region has approached or reached the critical tearing state in the forming simulation. Within this range, As a high-gain multiplier, it participates in the calculation of the coupling enhancement term and is included in the weighting coefficients. Under the influence of this, a significant nonlinear amplification effect is produced. Once the visual inspection layer captures the concurrent abrupt changes in reflection features and morphological features, the defect judgment index... It will quickly rise and exceed the preset alarm threshold, thereby enabling rapid identification of real stamping tear defects.

[0043] Furthermore, the three weighting coefficients mentioned above strictly satisfy the inequality constraints within the system: In a specific preferred embodiment, the system calibration value is... , , When only strong oil film reflection is observed in a certain area, the bottom adsorption and pressure holding are normal. , , , , The value is only 0.31, which is in a safe state; when a real metal scratch appears in a certain block, but has not yet cracked through the plate, , , , , The value jumped instantly to 2.21, exhibiting a highly distinctive abrupt change signal; when the deformation in a high-risk area had deteriorated into a penetrating tear, the underlying servo fan experienced a transient load decrease due to minor air leakage, triggering aerodynamic feedback that caused the vacuum leakage mutation factor to adaptively increase. Then the coupling enhancement term is amplified twice by the gaseous physical quantity. The value instantly jumped to 3.03, thus perfectly triggering the subsequent high-confidence interception and alarm commands.

[0044] When real metal tears or severe slip lines appear on the surface of automotive stamped parts, they not only manifest as simultaneous distortion of reflection and morphology, but this physical fracture inevitably occurs in the high thinning rate region predicted by CAE, introducing a vacuum leakage mutation factor. With FLD risk coefficient The joint product terms will be drastically activated, producing a significant nonlinear amplification effect; if it is a pool of drawn oil reflecting light in a flat area, due to the FLD risk coefficient of that area... Approaching zero, and the morphological anomaly index Extremely low, the entire coupling enhancement term is directly suppressed, completely eliminating misjudgment caused by stamping oil film.

[0045] Furthermore, to address the parameter compatibility issues arising from different materials and models of stamped parts during production line changes, this invention proposes a standardized offline parameter calibration and verification process. Before each batch of new model stamped parts is officially put into testing, the edge computing industrial control computer needs to calibrate the weighting coefficients using the following benchmark and extreme sample calibration method. , , To obtain the exact result: Step A: Extract 3 to 5 cleaned, uncoated, good-quality bare sheets as a baseline. The system is then activated to obtain the global average base reflectivity. Because highly reflective materials require higher optical tolerance, the system incorporates a negative correlation mapping relationship. The edge computing industrial control computer automatically calculates the basic reflection weight coefficient based on empirical constants. :when When approaching the extreme value of specular reflection, Approaching the lower limit of 0.1; when exhibiting diffuse reflection, Approaching the upper limit of 0.3. Subsequently, a good standard part coated with a standard oil film was introduced, and the extracted reflection interference index was systematically verified. and The product of these two terms is subject to a strict constraint that it must be strictly less than a safety baseline threshold when there is no abrupt change in appearance. 40%; Step B introduces a shape boundary limiter. At this point, aerodynamic feedback is normal and the FLD risk factor is extremely low. The system acquires the shape anomaly index. Then, the reverse solution is performed with the critical condition of just not triggering the second state alarm. The edge computing industrial control computer adjusts the basic weight coefficients of the topography. This makes it possible to achieve the following under the extreme scratch conditions: The calculation results reached This ensures that the system has the highest sensitivity to subtle morphological distortions while reserving a safety margin for process fluctuations; Step C: Introduce a sample with a through-and-through tear. The system simultaneously reads the abruptly changed vacuum leakage factor and the CAE simulation high-risk coefficient for that region. Using forced trigger state three as the target, the edge computing industrial control computer is set to a calibration target value of... The values ​​determined by substituting steps A and B , The nonlinear coupling gain coefficient is calculated in reverse. : ; At the same time, the system's underlying program will perform a verification: if the calculated... Inequality constraints are not satisfied If the calibration fails, the system will prompt the operator to re-prepare the torn sample or adjust the brightness of the underlying light source. Step D: After parameter calibration, the measurement system must be analyzed through a validation batch. The system needs to continuously perform mixed-flow testing on 50 good parts, 10 parts with oil film interference, and 5 parts with tear defects. The mandatory criteria for successful validation are: the false negative rate for tear defects must be absolutely 0%, and the recovery rate of parts with oil film interference that trigger state two and successfully reconstruct and release must be... Only when the verification criteria are met will the set of calibration parameters be packaged into a dedicated formula file for that model of stamped part and stored in the local database of control cabinet 30 for one-click retrieval during mass production.

[0046] Step S4, as follows Figure 7 As shown, the edge computing industrial control computer uses a defect judgment index. The system determines the spatial distribution of light sources and performs light field scheduling on the programmable light source device; the system presets a safety baseline threshold. With high-risk mutation threshold And satisfy ,in accordance with The numerical value strictly divides the detection block into three progressive processing states: when When the system is in state one, which is the safe release interval, it determines that the block is within the normal process fluctuation range in terms of both optical reflection and physical morphology, with no significant abnormalities. The edge computing industrial control computer releases the block directly without triggering any light source adjustment or re-shooting actions, so as to maximize the overall inspection cycle of the production line.

[0047] when In state two, the intermediate warning zone, the system determines that there is a high probability of optical reflection artifacts caused by the stretching oil film in the area. Relying on the vertical high-precision fine-tuning slide 29 mounted above the light-transmitting hole of the hemispherical light source cover 18, the main optical axis of the industrial camera 26 is ensured to pass through the geometric center of the hemispherical dome. This mechanical rigid constraint ensures that the spatial coordinates of the visual target surface achieve high-precision spatial position consistency in the two physical exposures before and after the polarization light switching.

[0048] The specific reshooting scheduling process is as follows: First, based on the spatial coordinates of the block in the image, the corresponding light source node combination is determined in a preset three-dimensional light field topology mapping table. This mapping table is pre-constructed by establishing a precise relationship matrix between the physical coordinates of the light source and the three-dimensional spatial illumination area during the system calibration phase. The three-dimensional light field topology mapping table in the edge control center is generated by the edge computing industrial control computer through offline CAE pre-mapping and online physical calibration fusion. The specific construction and update steps are as follows: First, the system imports the 3D CAD model of the current batch of stamped parts and extracts the local mesh blocks. The three-dimensional normal vectors are used to calculate the set of light source nodes that theoretically cause glare artifacts in the block using a three-dimensional ray tracing algorithm. Subsequently, standard good parts free of oil are introduced into the production line. The control unit sequentially illuminates the visible light LED array 19 according to spatial sectors and simultaneously performs high-speed continuous shooting to obtain the actual light source and spatial illumination influence matrix. The edge computing industrial control computer orthogonally merges the theoretical calculation results with the actual physical matrix to generate a topology mapping table specific to this model. This mapping table is based on grid blocks. Using the spatial coordinates as an index, three sets of light source control commands are bound to it: normal lighting group, glare suppression group, and polarization substitution group. When switching stamping part models, the system automatically retrieves the new CAD model according to the model change command issued by the MES system, and performs a quick LED sector traversal calibration when the first part passes through the vision station. This automatically overwrites and updates the three-dimensional light field topology mapping table, realizing flexible line change without manual intervention.

[0049] Next, the system shuts down the specific visible light LED nodes that generate glare and simultaneously turns on the corresponding polarized light LED array 20 to change the local illumination angle and polarization direction. Then, the industrial camera 26 re-acquires an image of the area at the original physical location and feeds the re-enhanced image back to step S2 for secondary feature analysis. The light field distribution is adjusted in real time based on the visual analysis results, enabling the detection system to form an adaptive light field closed-loop control mechanism based on the defect probability distribution. Through this physical interaction process of image analysis, light field reconstruction, and re-enhancing verification, reflective interference is gradually stripped away and eliminated until the area is recalculated. The value either falls back to state one or jumps to state three; In the optical field reconstruction closed loop of state two described above, the specific adjustment parameters of the polarized LED array 20 are achieved based on PWM high-frequency duty cycle dimming. The edge computing industrial control computer determines the reflection interference index of the current block. The overshoot is linearly mapped to the output PWM duty cycle, adjusting the polarized light intensity with a millisecond-level response. Simultaneously, the polarization angle is controlled by an embedded liquid crystal tunable polarizer. to The angle adjustment resolution of the liquid crystal tunable polarizer is preferably finely adjusted electronically. Response time of electronic switching between wavelength and polarization state To match the optimal extinction angle of the surface normal. The convergence criterion for re-scanning is: the absolute difference between the defect judgment index calculated by two consecutive in-situ re-scanning iterations. Or, the maximum number of anti-dead-loop reconstructions set by the system is reached. If it reaches back If it is still in state two, it will be forced to transition to state three and be removed.

[0050] when When the system is in state three, i.e., the high-risk defect interval, it determines that the reflection and morphology distortion of this block have been affected by the core variables. The coupling terms are amplified dramatically, indicating a real physical defect with extremely high confidence. The system triggers a fidelity interception mechanism, directly freezing the light field reconstruction operation in that area, breaking out of the closed loop, and sending the coordinates and type information of the high-risk defect directly to the alarm output unit.

[0051] Preferably, the security baseline threshold With high-risk mutation threshold It is not a fixed constant, but a dynamic threshold adaptively generated by the edge computing industrial control computer during the system calibration phase; the system first determines the safety baseline threshold based on the basic surface roughness information of the current batch of automotive stamping parts. This threshold is used to characterize the lower limit of the defect judgment index under normal surface conditions; at the same time, the high-risk mutation threshold is determined in conjunction with the CAE forming limit constraint conditions of automotive stamping parts. This threshold is used to define the upper limit of the alarm for potential cracking or tearing risks. Based on this, the edge computing industrial control computer performs adaptive segmentation calculation on the detection data through the maximum inter-class variance method, thereby generating a dynamic threshold suitable for the current batch of inspections. Through the above design, the closed-loop control mechanism can automatically adapt to the inspection conditions under different sheet materials and drawing oil coating thicknesses. Furthermore, the security baseline threshold With high-risk mutation threshold The adaptive generation mechanism is achieved by the edge computing industrial control computer through steps of anomaly cleaning, self-tuning, and hysteresis closed-loop: First, before feature extraction, the system uses median filtering to remove sensor defects and uses the isolated forest algorithm to isolate extreme ambient light disturbances and flying debris noise in the workshop, ensuring the authenticity of the input data and solving the threshold drift problem caused by equipment aging and environmental interference. Secondly, the system utilizes historical sliding windows to collect effective mesh features of the reference stamping parts and executes the maximum inter-class variance method to obtain the basic segmentation operator. Based on the average cracking risk coefficient of the key characteristic areas of this batch Perform self-tuning solution: ; ; in: and b represents the material reflectance distribution coefficient; b represents the noise suppression coefficient. c represents the mean of the background noise; c represents the relaxation penalty factor. This represents the physical safety limit value; the formula implements strong constraints on the pure visual segmentation operator based on physical and mechanical properties. The system prevents defect judgment index The system induces high-frequency jitter in the optical field at the threshold critical point, and sets a bidirectional limiting hysteresis bandwidth. When a detection block triggers a state two alarm, the release condition for reverting to state one is strictly limited to: This completely eliminates state oscillations and ensures the stability of optical field scheduling.

[0052] Step S5: After the above dynamic light field closed-loop control reaches the convergence state, the alarm output unit will output the signal based on the final stable state. The probability distribution results accurately output the spatial location, area range, and specific defect type information of the surface defects of the stamped parts, and send the detection results to the production line control system to complete the surface defect detection process of the stamped parts. Based on the final converged defect judgment index state in step S4, the edge computing industrial control computer generates corresponding classification and diversion control commands: If it is determined that all detection blocks of the current stamped part are ultimately in state one, the edge computing industrial control computer does not output a rejection signal, and the workpiece continues to be smoothly conveyed along the positive X-axis by the flexible negative pressure conveyor belt 7, entering the next production or packaging process; If it is determined that any block of the current stamped part is ultimately in state three, the edge computing industrial control computer immediately extracts the pulse count value of the absolute encoder 14 at the moment the workpiece is triggered for imaging at the vision inspection station. Since there is a fixed known physical distance between the center of the main optical axis of the industrial camera 26 at the vision inspection station and the center of the push rod of the defect rejection device in the X-axis direction, the edge computing industrial control computer converts it into the corresponding rejection delay pulse number; The system background monitors and accumulates the displacement pulse signals output by the absolute encoder 14 in real time for kinematic tracking. When the accumulated pulse value reaches the rejection delay pulse count, it is determined that the defective workpiece has been accurately delivered to the defective rejection station. At this time, the edge computing industrial control computer immediately outputs a high-frequency trigger signal to the defective rejection device to control the opening and closing response time. The high-speed solenoid valve is instantly activated, and the single extension execution time of the pneumatic push rod 16 is set to... nominal lateral thrust Thus, in the high-speed motion state of the production line without stopping, the defective workpiece is neatly pushed out of the flexible negative pressure conveyor belt 7 and falls into the waste collection tank 17 on the side. Subsequently, the pneumatic push rod 16 automatically retracts and resets, completing the physical sorting loop for defective products. Simultaneously, the edge computing industrial control computer uploads the defect type, coordinate location, and original image data of the defective product to the local database for subsequent quality traceability and process optimization.

[0053] Example 4, as Figure 8 As shown, to further verify the feasibility and robustness of the above mechanical architecture and coupled control algorithm in a real industrial environment, this embodiment provides a specific application scenario example; The test subject was a high-reflectivity aluminum alloy front door outer panel of a certain vehicle model. The stamped part was coated with drawing oil and included a flat main surface of the door panel and a door handle recess with a large drawing depth. The system's preset threshold parameter is: safety baseline threshold. High-risk mutation threshold The algorithm weights are calibrated as follows: , , ; The system encountered three typical detection conditions simultaneously during operation. Conditions A and B involved surface defects or artifacts, with no penetrating leaks. The pressure in the bottom vacuum adsorption box 11 was good, and the servo fan reported a sudden change in vacuum leakage factor. The value was maintained at the baseline of 1.0; however, condition C was a severe through-tear, resulting in transient micro-leakage, and the pneumatic feedback caused... The value was increased to 1.5. Its visual extraction parameters and the retrieved prior parameters are shown in Table 1. Table 1 Typical Detection Condition Parameter Settings

[0054] Among them: reflection interference index Characterizing the intensity of surface reflective disturbance, morphological anomaly index Characterizing the degree of local geometric discontinuity detected by vision, the FLD risk coefficient Characterizes the degree of danger of this region on the forming limit diagram; The edge computing industrial control computer performs real-time calculations on the above three working conditions based on the FLD modulation coupling damage index model. The calculation results and judgment status are shown in Table 2.

[0055] Table 2 Defect Judgment Index Calculation results

[0056]

[0057] According to the system's preset judgment rules: when When the system determines the status to be state one, it allows passage directly; when When the system determines it to be in state two, it initiates light field reconstruction and in-situ re-shooting; when When the system determines the state to be three, it will directly output a high-confidence defect alarm. For the above-mentioned working condition B, the system determines that it has entered state two. The light source control unit initiates the light field reconstruction process, turns off the visible light LED node that generates glare, turns on the polarized light LED array 20 at the corresponding position, and performs secondary feature analysis at the original physical position; the index comparison before and after the reshoot is shown in Table 3:

[0058] Table 3. Comparison of DI before and after reshooting the light field reconstruction in State 2.

[0059]

[0060] Since the dust at this location is a pure optical artifact, after polarized light is used to eliminate diffuse reflection and glare from the surface, the morphological anomaly index is extracted a second time. A significant decrease occurred. (Recalculated) The value dropped to 0.348, successfully returning to the safe release range. Based on this, the system investigated the truth and lifted the alarm, effectively avoiding unnecessary downtime caused by false alarms.

[0061] To evaluate the overall reliability of the device and algorithm, continuous testing and verification were conducted on 1000 door outer panels from the same batch in a real stamping workshop environment. The statistical results after manual verification are shown in Table 4.

[0062] Table 4. Statistical Analysis of System False Alarm Rate and False Negative Alarm Rate

[0063]

[0064] Experimental results show that by introducing a nonlinear coupling criterion between the FLD risk coefficient and optical features, and combining a closed-loop re-shooting mechanism in state two, the present invention can effectively suppress false alarms caused by oil film reflection and dust interference, while ensuring high sensitivity and zero false alarms for real crack defects. As can be seen from the above three typical working condition examples and batch inspection statistics, the FLD modulation coupling damage index model proposed in this invention can effectively distinguish between optical artifacts and real structural cracks in a high-reflectivity drawing oil environment. When the detection area is in the safe forming zone, even if the reflection interference index is high, the system can still significantly weaken the coupling enhancement term through the suppression effect of the FLD risk coefficient, thereby avoiding false alarms. In the high-risk area of ​​deep drawing, small morphological anomalies will not only be amplified by the mechanical nonlinear coupling term, but more importantly, when the defect deteriorates into a through-thaw tear, the unique aerodynamic leakage factor of this system will help to mitigate the damage. It will be activated as a quadratic gain multiplier, thus making the final defect determination index... A dramatic exponential leap occurs, thus providing an extremely reliable cross-physics cross-verification defense against malignant defects.

[0065] In summary, this invention not only improves the accuracy of surface defect detection in stamped parts, but also significantly enhances the robustness and reliability of opto-mechatronic systems in complex industrial environments.

[0066] It should be noted that, for the sake of brevity, the foregoing method embodiments are described as a series of actions, but this does not mean that the application limits the order of the steps. Based on the ideas of this application, some steps can be executed in different orders or in parallel without affecting the functional implementation. Secondly, those skilled in the art should also understand that the specific embodiments described in the specification are preferred embodiments of the technical solutions of this application, and not limitations on the scope of protection of this application. All equivalent improvements or substitutions made within the spirit and principles of this application should be covered within the scope of protection of this application.

[0067] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vision analysis-based inspection device for automotive stamped parts, characterized in that, The device includes a frame, a workpiece conveying device, a programmable light source device, a vision inspection device, a defective product rejection device, and a control cabinet. The workpiece conveying device is horizontally mounted on the lower middle part of the frame, penetrating the detection area, and includes a servo frequency converter fan. The programmable light source device is suspended directly above the detection area; The visual inspection device adopts a zenith coaxial vertical top-view layout and includes an industrial camera. The defective product rejection device is located at the discharge end of the workpiece conveying device along the workpiece conveying direction; The control cabinet contains an edge computing industrial control computer. Based on the above structure, the edge computing industrial control computer is used to receive image data acquired by the industrial camera and divide it into local grid blocks to obtain the reflection interference index and morphology anomaly index of each block; the edge computing industrial control computer is also used to determine the vacuum leakage mutation factor according to the load current signal of the servo variable frequency fan drive circuit, and extract the critical cracking risk coefficient of the corresponding block according to the offline CAE forming limit diagram; the edge computing industrial control computer constructs the FLD modulation coupling damage index model, and inputs the reflection interference index, the morphology anomaly index, the vacuum leakage mutation factor and the critical cracking risk coefficient as detection parameters into the FLD modulation coupling damage index model to obtain the defect judgment index; the edge computing industrial control computer will perform dynamic light field scheduling control on the programmable light source device to realize light field reconstruction according to the spatial distribution state of the defect judgment index, and perform in-situ re-shooting through the industrial camera to obtain the reconstructed light field image for secondary analysis, and link the defect rejection device for physical sorting when a defect is found.

2. The automotive stamping parts inspection device based on vision analysis according to claim 1, characterized in that, The frame includes: a base beam, damping and shock-absorbing feet, a gantry column, and a top support platform; Multiple base beams are orthogonally welded to each other to form a rectangular chassis structure, and the damping and shock-absorbing feet are screwed to the four corners of the bottom. The gantry column extends vertically upward from the end of the base beam and is rigidly connected at the top via the top support platform.

3. The automotive stamping parts inspection device based on vision analysis according to claim 1, characterized in that, The workpiece conveying device also includes: a flexible negative pressure conveyor belt, a servo drive motor, an active roller, a driven tensioning roller, a vacuum adsorption box, an absolute encoder, and a laser beam sensor. The axes of the driving roller and the driven tensioning roller are both parallel to the horizontal plane and perpendicular to the workpiece conveying direction. The servo drive motor is coaxially connected to one end of the driving roller, and the absolute encoder is coaxially connected to the other end of the driving roller. The laser beam sensor spans both sides of the conveying path at the feeding end.

4. The automotive stamping parts inspection device based on vision analysis according to claim 3, characterized in that, The vacuum adsorption box is embedded between the active roller and the driven tensioning roller. Its upper end face is open and close to the bottom surface of the flexible negative pressure conveyor belt. Its lower end is sealed and connected to the servo frequency conversion fan at the bottom through a flexible air pipe. The surface of the flexible negative pressure conveyor belt is coated with a soft coating with a high coefficient of friction, and the surface of the flexible negative pressure conveyor belt is uniformly distributed with an array of micropores.

5. The automotive stamping parts inspection device based on vision analysis according to claim 4, characterized in that, The upper surface of the vacuum adsorption box is provided with a low-friction support plate with a flow-guiding grid groove; A gas-liquid separation filter is connected in series between the flexible air pipe at the bottom of the vacuum adsorption box and the servo frequency converter fan; a current sampling module is set in the drive circuit of the servo frequency converter fan to collect the fan load current signal in real time and upload it to the edge computing industrial control computer.

6. The automotive stamping parts inspection device based on vision analysis according to claim 1, characterized in that, The programmable light source device specifically includes: a hemispherical light source cover, a visible light LED array, a polarized light LED array, an ultraviolet LED array, a heat dissipation conformal plate, an adjustment frame, adjustment bolts, and a light source control circuit board; The visible light LED array, the polarized light LED array, and the ultraviolet LED array are arranged in concentric circles, nested alternately, on the concave hemispherical surface of the hemispherical light source cover; the heat dissipation conformal plate is tightly fitted to the convex spherical surface of the hemispherical light source cover.

7. The automotive stamping parts inspection device based on vision analysis according to claim 6, characterized in that, One end of the adjustment frame is fixed to the top support platform, and the other end is connected to the hemispherical light source cover. The adjustment frame is equipped with the adjustment bolt, which allows the hemispherical light source cover to be finely adjusted in the vertical direction.

8. The automotive stamping parts inspection device based on vision analysis according to claim 6, characterized in that, The hemispherical light source cover has a light-transmitting hole at its geometric apex, and the visual inspection device also includes: a large depth-of-field dual telecentric lens, a vertical high-precision micro-adjustment slide, and a polarizing filter. The vertical high-precision fine-tuning slide is vertically mounted on the outside above the light-transmitting hole; after the industrial camera and the large depth-of-field double telecentric lens are assembled, they are mounted on the vertical high-precision fine-tuning slide, and the polarizing filter is screwed to the front end of the lens; the main optical axis of the industrial camera is strictly parallel to the vertical direction and passes through the center of the hemisphere.

9. The automotive stamping parts inspection device based on vision analysis according to claim 1, characterized in that, The control cabinet is independently mounted to isolate the production line from vibration, and it also includes a high-voltage distribution module. The defective product rejection device is located on the side of the discharge end and includes a high-speed solenoid valve and a pneumatic push rod. Its control end is electrically connected to the edge computing industrial control computer and is used to push out defective stamped parts laterally.

10. The automotive stamping parts inspection device based on vision analysis according to claim 1, characterized in that, The control cabinet integrates an edge control hub, which includes: The image acquisition unit is used to receive the initial light field image of the stamped part surface acquired by the industrial camera, as well as the reconstructed light field image captured in situ after the light field closed-loop adjustment. The feature analysis unit is used to divide the acquired initial light field image or reconstructed light field image into local grid blocks, solve the optical and geometric properties of each block in parallel, and output the reflection interference index and the morphology anomaly index. The coupling decision unit is used to construct the FLD modulation coupling damage index model, retrieve the offline CAE forming limit map to extract the critical cracking risk coefficient of the corresponding block, and receive the vacuum leakage mutation factor calculated by the edge computing industrial control computer based on the servo variable frequency fan load current signal. The reflection interference index, the morphological anomaly index, the critical cracking risk coefficient and the vacuum leakage mutation factor are jointly coupled and calculated to obtain the defect judgment index, thereby generating a global defect probability distribution field. The light source control unit has an embedded topology mapping table, which is used to receive the global defect probability distribution field signal output by the coupled decision unit, and to start the dynamic light field adaptive control model when it is determined that light field optimization is required, and output spatial coordinate control commands to the light source control circuit board to achieve light field reconstruction by adjusting the spatial illumination distribution of the programmable light source device in a closed loop. The alarm output unit is used to output the defect coordinates and defect type with high confidence after the optical field closed-loop adjustment is completed.

Images