Automobile stamping quality sorting method and system combined with optical detection

Abstract

This invention relates to the interdisciplinary field of mechanical engineering and optical inspection, and discloses a method and system for quality sorting of automotive stamping parts combined with optical inspection. The method includes: synchronous imaging by a high-speed linear array camera and a structured light projector via an encoder and photoelectric triggering device; fusing 3D point clouds and 2D images, and using a deep convolutional neural network for defect identification and quality grading; calculating the six-degree-of-freedom pose of the workpiece using an iterative nearest-point algorithm and principal component analysis; calculating the spatiotemporal parameters of sorting based on a predictive instruction generation module, and issuing dynamic control commands to multi-degree-of-freedom pneumatic push rods via a hard real-time communication bus to execute sorting. The system includes a synchronous triggering unit, an imaging unit, a defect identification unit, a pose calculation unit, a predictive instruction generation module, and a dynamic sorting execution unit. This invention achieves millisecond-level accurate identification and adaptive sorting, significantly improving detection accuracy, sorting efficiency, and system robustness.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of mechanical engineering and optical inspection, specifically to a method and system for quality sorting of automotive stamping parts that combines optical inspection. Background Technology

[0002] In modern automotive manufacturing, stamping is the primary forming method for body panels, and its production quality directly impacts the assembly precision and efficiency of the entire vehicle. With continuously increasing automotive production capacity, the cycle time requirement for stamping production lines has generally exceeded 30 pieces / minute. Traditional sorting methods relying on manual visual inspection or semi-automatic equipment are inefficient, subjective, and inconsistent, making them unsuitable for the quality control demands of high-speed continuous production. To improve automation, automated optical inspection (AOI) systems based on machine vision are gradually being introduced to the end of stamping production lines, attempting to achieve defect identification and classification through image comparison or 3D reconstruction.

[0003] However, existing methods have significant shortcomings in real-world high-speed production environments: First, most vision systems use a fixed frame rate for shooting, which is not synchronized with the conveyor belt movement. This makes them prone to motion blur or image misalignment due to speed fluctuations, leading to missed defects. Second, the system typically only outputs the two-dimensional center coordinates of the workpiece, lacking the ability to perceive three-dimensional postures (such as rotation and warping) caused by stamping springback, stacking, or conveyor offset, resulting in inaccurate positioning of subsequent actuators or even collisions. Third, the link delay from image acquisition to sorting execution is large, usually exceeding 100 milliseconds. Under high-speed cycles, it is difficult to match the command with the workpiece position, causing false rejections or missed sorting. Moreover, existing sorting execution strategies are mostly statically preset and cannot dynamically adjust the point of application and force according to the real-time posture of the workpiece, resulting in poor adaptability to posture changes.

[0004] While existing research has incorporated deep learning to improve defect identification accuracy, it has not yet effectively addressed core issues such as imaging synchronization, pose integrity perception, system response latency, and execution adaptability. Therefore, there is an urgent need for an integrated technical solution capable of achieving precise synchronous imaging, real-time six-DOF pose calculation, predictive instruction generation, and dynamic adaptive sorting under high-speed operating conditions, in order to overcome the current comprehensive bottlenecks in accuracy, speed, and robustness of sorting systems. Summary of the Invention

[0005] This invention provides a quality sorting method and system for automotive stamping parts that combines optical inspection, aiming to solve the technical problems in the prior art, such as asynchronous image acquisition and production line movement, insufficient pose estimation capability, disconnect between detection and execution, and low detection accuracy, high execution delay, and poor sorting success rate caused by static sorting strategies.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for quality sorting of automotive stamping parts incorporating optical inspection is provided, comprising the following steps: An encoder and a photoelectric triggering device are installed on the stamping part conveying path. The encoder is installed on the conveyor belt drive shaft and is used to output the conveyor belt running speed signal in real time. The photoelectric triggering device is arranged in the workpiece inlet area along the conveying direction and is used to detect the position of the workpiece leading edge and generate an initial trigger signal. Based on the initial trigger signal and the speed signal output by the encoder, the time window for the workpiece to reach the preset imaging area is calculated, and a synchronous trigger command is sent to the high-speed linear scan camera and the structured light projector accordingly. This causes the high-speed linear scan camera to perform line-by-line exposure when the workpiece passes through the imaging area in a way that is relatively stationary with respect to the workpiece surface, while the structured light projector projects a multi-frequency phase-shifting stripe pattern onto the workpiece surface. The high-speed linear array camera acquires a sequence of structured light images reflected from the workpiece surface, and the three-dimensional point cloud data of the workpiece surface is calculated based on the phase unfolding algorithm. At the same time, a deep convolutional neural network is used to fuse the three-dimensional point cloud data with the corresponding two-dimensional grayscale image to extract the defect features of the workpiece surface and determine the quality level of the workpiece. Based on the three-dimensional point cloud data, the iterative nearest point algorithm combined with principal component analysis is used to calculate the six-degree-of-freedom pose parameters of the workpiece in space, including the translation along the X, Y, and Z axes and the rotation angles around the X, Y, and Z axes. The quality level determination result and the six-degree-of-freedom pose parameters are input into the predictive instruction generation module. This module calculates the target action time and spatial action point of the sorting actuator based on the real-time position of the current workpiece on the conveyor belt, the conveying speed, and the physical response delay of the downstream sorting actuator. The target action time and spatial point of action are directly sent to the servo controller of the sorting execution mechanism through a hard real-time communication bus. The servo controller drives the multi-degree-of-freedom pneumatic push rod to act on the specified position of the workpiece at a preset time with a specified force and direction, so as to complete the dynamic sorting of qualified products and defective products.

[0007] In one embodiment of the present invention, the line frequency f of the high-speed linear array camera is determined by the following formula:

[0008] in The conveyor belt speed is measured in real time by the encoder. The pixel physical size of the high-speed linear array camera is the ground sampling distance determined by the camera lens focal length, sensor pixel size, and object distance, ensuring that the image is free of motion blur.

[0009] In one embodiment of the present invention, the multi-frequency phase-shifted fringe pattern projected by the structured light projector includes three sets of sinusoidal fringes with different spatial frequencies, each set containing four fringes with a phase difference of [missing information]. The image is used to calculate the wrapping phase using the four-step phase shift method, and combined with the three-frequency heterodyne method for phase expansion to obtain the absolute phase map.

[0010] As one embodiment of the present invention, the deep convolutional neural network is a two-stream encoder-decoder architecture, wherein the first stream input is a two-dimensional grayscale image, which extracts texture and edge features through five convolutional and max pooling operations; the second stream input is three-dimensional point cloud data, which is voxelized and then extracted through a three-dimensional convolutional layer to extract geometric features; the two stream features are concatenated at the bottleneck layer and weighted and integrated through an attention fusion module, and finally outputs a pixel-level defect classification map, the defect types of which include scratches, dents, wrinkles, oil stains and edge defects.

[0011] As one embodiment of the present invention, the six-degree-of-freedom pose calculation process includes: firstly, segmenting the workpiece body region from the three-dimensional point cloud and removing the background and supporting structure; denoising and normal vector estimation of the segmented point cloud; then, using the standard CAD model point cloud as a reference, performing coarse registration through an iterative nearest-point algorithm; based on the coarse registration, employing a local optimization method based on principal component analysis to compensate for the slight warping of the workpiece caused by stamping springback, and outputting accurate six-degree-of-freedom pose parameters.

[0012] As one embodiment of the present invention, the predictive instruction generation module incorporates a kinematic model of the sorting actuator, which includes the complete time delay from receiving an instruction to completing the action. , The response time of the servo valve, the acceleration time of the cylinder, and the mechanical transmission backlash are all determined by the servo valve response time, the cylinder acceleration time, and the mechanical transmission backlash. The predictive command generation module calculates the position of the workpiece based on its current velocity v and pose. Position after time The predicted position is combined with the workpiece posture and mapped to the spatial action point of the sorting execution mechanism in the workpiece body coordinate system, where P_current is the current position of the workpiece in the world coordinate system calculated based on the same encoder signal, and v is the instantaneous speed of the conveyor belt obtained in real time by the encoder.

[0013] According to another aspect of the present invention, a quality sorting system for automotive stamping parts incorporating optical detection is provided, comprising: The synchronous triggering unit generates synchronous triggering commands for the camera and the structured light projector based on photoelectric triggering signals and encoder speed signals; The imaging unit includes a high-speed linear array camera and a multi-frequency structured light projector, used to acquire two-dimensional images and three-dimensional topographic data of the workpiece surface; The defect identification and quality assessment unit is used to detect and classify defects in the fused multimodal data through a deep convolutional neural network and output a quality level. A six-degree-of-freedom pose calculation unit calculates the complete pose parameters of a workpiece in space based on three-dimensional point cloud data; The predictive instruction generation module is used to calculate the target spatiotemporal parameters of the sorting action by combining pose parameters, transmission speed and actuator dynamics. The dynamic sorting execution unit includes a multi-degree-of-freedom pneumatic push rod array and its servo controller, which is used to perform sorting actions with specified parameters at a specified time; A hard real-time communication bus connects the predictive instruction generation module and the dynamic sorting execution unit, with a communication cycle of no more than 1 millisecond.

[0014] As one embodiment of the present invention, the multi-degree-of-freedom push rod array includes three independently controlled pneumatic push rod units arranged along the conveying direction; each pneumatic push rod unit is equipped with a proportional servo valve and a displacement sensor, which can independently control the extension length, the force applied and the triggering sequence.

[0015] In one embodiment of the present invention, the hard real-time communication bus is an EtherCAT bus, the master station cycle is set to 500 microseconds, and distributed clock synchronization is supported, and the slave station is the servo controller.

[0016] As one embodiment of the present invention, the synchronous triggering unit is integrated into an industrial field programmable gate array chip. After receiving the initial trigger signal, its internal logic circuit immediately latches the current encoder count value and, based on the preset imaging distance and transmission speed, outputs a trigger pulse to the high-speed linear array camera and structured light projector at the calculated precise time through a hardware timer. The output trigger pulse jitter is less than 5 microseconds.

[0017] In one embodiment of the present invention, the six-degree-of-freedom pose calculation unit and the defect identification and quality judgment unit share the same three-dimensional point cloud input, the point cloud resolution is not less than 8 points per square centimeter, and the point cloud registration error is less than 0.1 mm.

[0018] In one embodiment of the present invention, the system further includes a workpiece positioning confirmation module, which consists of a secondary photoelectric switch located downstream of the imaging area. This module verifies whether the workpiece passes through the imaging area as expected and initiates a compensation imaging process in case of missed detection or false triggering. The compensation imaging process is executed according to the following steps: When the secondary photoelectric switch fails to detect a workpiece within the expected time window, it immediately sends a pause signal to the PLC to stop the conveyor belt. The auxiliary area array camera located above the imaging area is activated to take a high-resolution re-image of the stationary workpiece. Simultaneously, a structured light projector projects fast single-frequency fringes to acquire a 3D point cloud. The re-image (2D) and the 3D point cloud are then sent to the defect identification unit and pose calculation unit for re-assessment and re-calculation of quality and pose. The recalculated results are input into the predictive instruction generation module to execute the sorting action. After completion, the PLC resumes conveyor belt operation.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: By fusing signals from the encoder and photoelectric triggering device, and dynamically matching the line scan camera's frequency with the conveyor belt speed, the image acquisition process is synchronized with the workpiece movement at the hardware level. This mechanism ensures, in principle, that even under continuous high-speed conveying conditions, workpiece images with no motion blur and precise positioning can still be acquired, providing a reliable data foundation for all subsequent processing stages. This innovative approach combines structured light 3D reconstruction with deep learning image recognition. On one hand, it quantifies the geometric shape deviations of the workpiece using 3D point clouds, effectively detecting 3D defects such as dents and warping. On the other hand, it accurately identifies surface imperfections such as scratches and oil stains by fusing 2D image texture features. Furthermore, it calculates complete six-degree-of-freedom pose based on the fused data, overcoming the limitation of existing technologies that only provide 2D planar coordinates, and accurately sensing spatial pose changes of the workpiece caused by springback and stacking. By introducing a fixed response delay of the sorting actuator into the built-in kinematic model Based on real-time acquired workpiece speed and pose, feedforward calculations are performed to predict and plan the timing and point of action for sorting actions in advance. This mechanism innovates the traditional serial "detection-processing-triggering" mode into a parallel "detection-prediction-cooperative execution" mode, effectively compensating for the end-to-end latency from information perception to physical execution. Based on the precise workpiece body coordinate system pose, the optimal sorting action point and thrust direction are dynamically calculated and executed through independent closed-loop control of a multi-degree-of-freedom pneumatic pusher array. This strategy enables the sorting action to adaptively cope with the random translation, rotation, and warping of the workpiece, avoiding missed pushes or collisions under static strategies, and significantly improving the robustness of the sorting system to complex working conditions. By employing a hard real-time industrial communication bus to directly connect the decision-making unit and the execution unit, replacing the traditional PLC polling relay mode, predictive control commands can be issued and executed with extremely low deterministic latency. This architecture, combined with the predictive command generation module, forms a fast-responding and timing-precise control closed loop, meeting the stringent requirements of high-speed production cycles. Attached Figure Description

[0020] Figure 1 This is a structural block diagram of the sorting system in this invention; Figure 2 This is a flowchart of the sorting method in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention in any way. Example 1

[0022] This invention provides a method and system for quality sorting of automotive stamped parts combined with optical detection. An encoder and a photoelectric triggering device are installed along the stamped part conveyor path. The encoder is mounted on the conveyor belt drive shaft and outputs the conveyor belt speed signal in real time. The photoelectric triggering device is arranged along the conveying direction in the workpiece entry area and detects the leading edge position of the workpiece to generate an initial trigger signal. The encoder is an incremental rotary encoder with a resolution of 10,000 pulses per revolution. It outputs A and B phase signals through an orthogonal decoding circuit, and the conveyor belt linear speed is calculated in real time by an internal counter of a field-programmable gate array (FPGA) chip. The photoelectric triggering device uses a diffuse reflection infrared photoelectric switch with a response time of less than 100 microseconds. Its sensing surface is perpendicular to the conveying direction. When the leading edge of the stamped part enters the sensing area, a rising edge level signal is immediately generated as the initial trigger signal. This signal is synchronously latched with the current encoder count value, forming a time reference for subsequent imaging triggering.

[0023] Based on the initial trigger signal and the speed signal output by the encoder, the time window for the workpiece to reach the preset imaging area is calculated, and a synchronous trigger command is sent to the high-speed line scan camera and the structured light projector accordingly. The preset imaging area is located in the middle of the conveyor path, and the physical distance from the photoelectric trigger device is L, in millimeters. The current instantaneous speed v of the conveyor belt is calculated from the encoder pulse frequency, in millimeters per second. Therefore, the time required for the leading edge of the workpiece to reach the imaging area is t_arrival = L / v. The synchronous trigger command for the high-speed line scan camera and the structured light projector is issued Δt before t_arrival, where Δt is the sum of the start-up delays of the camera and the projector, which is taken as 800 microseconds in this embodiment. This trigger command is generated by the internal hardware timer of the field-programmable gate array chip, outputting a standard TTL level signal with trigger jitter of less than five microseconds, ensuring strict alignment between the imaging time and the workpiece position. The formula for calculating the line frequency f of the high-speed line scan camera is: ,in The conveyor belt speed is measured in real time by the encoder. The pixel physical size of the high-speed linear array camera is determined by the camera lens focal length, sensor pixel size, and object distance. In this embodiment, the value is 0.05 mm, corresponding to a conveyor belt speed of 1.2 m / s and a line frequency of 24 kHz, to ensure that the workpiece surface is relatively stationary during each line of exposure in the imaging process, thus completely eliminating motion blur.

[0024] The high-speed linear array camera acquires a sequence of structured light images reflected from the workpiece surface, and the three-dimensional point cloud data of the workpiece surface is calculated based on a phase unfolding algorithm. A structured light projector projects three sets of sinusoidal fringe patterns with different spatial frequencies. Each set contains four images with a phase difference of π / 2, denoted as follows: , , , For any frequency f, the wrapping phase φ_wrapped is calculated using a four-step phase-shift method: .

[0025] The three sets of frequencies are respectively =16 cycles per image, =8 cycles per image, =4 periods per image, phase unwrapping is performed using the three-frequency heterodyne method to obtain an absolute phase map φ_absolute. The absolute phase map is mapped to the world coordinate system using calibration parameters to generate 3D point cloud data. The point cloud resolution is no less than eight points per square centimeter, with uniform point spacing, covering the entire surface of the stamped part, including edges and curved areas.

[0026] Simultaneously, a deep convolutional neural network is used to fuse the 3D point cloud data with the corresponding 2D grayscale image to extract surface defect features of the workpiece and determine its quality level. The 2D grayscale image is synchronously acquired by a high-speed linear scan camera under unstructured light illumination conditions and is strictly spatiotemporally aligned with the 3D point cloud. The deep convolutional neural network adopts a two-stream encoder-decoder architecture. The first stream input is a 2D grayscale image with a size of 2048×1536 pixels, which is sequentially processed through five layers of convolution and max pooling operations. Each convolutional kernel has a size of 3×3 and the number of channels is 64, 128, 256, 512, and 1024, respectively, to extract texture, edge, and local contrast features. The second stream input is 3D point cloud data, which is first voxelized with a voxel side length of 1 mm to form a 128×128×64 3D mesh, and then processed through four 3D convolutional layers with a kernel size of 3×3×3 and the number of channels is 32, 64, 128, and 256, respectively, to extract geometric shape, curvature changes, and concavity / convexity anomaly features. The two-stream features are concatenated at the bottleneck layer to form a fused feature tensor, which is then input into the attention fusion module. This module calculates the spatial correlation weights between the 2D texture features and the 3D geometric features, weighting and enhancing high-confidence regions to suppress noise interference. The attention fusion module employs a cross-attention mechanism, where the query comes from the 2D texture feature map, and the key and value come from the 3D geometric feature map; specifically, let the 2D feature be... The three-dimensional voxel features are First Average pooling in the depth dimension yields Then, key and value are generated through linear projection; attention weights. The fusion feature is Where H and W represent the spatial height and width of the feature map, respectively. and These represent the number of channels in the two-dimensional feature flow and the three-dimensional feature flow, respectively. Q represents the number of channels in the two-dimensional feature flow. The query matrix generated by linear projection Key matrix The dimension of each key vector in the matrix is ​​used to scale the dot product attention, and its value is equal to... Or the channel dimension after projection, This represents the calculated attention weight matrix. T represents the feature tensor output after attention-weighted fusion, where T is a matrix. The transpose operator is used. Finally, upsampling and convolution are used to output a pixel-level defect classification map. Defect types include five categories: scratches, dents, wrinkles, oil stains, and edge defects, each assigned a unique integer label. If the defect area exceeds a preset threshold or is located in a critical functional area, it is judged as a defective product; otherwise, it is a qualified product.

[0027] Based on the aforementioned 3D point cloud data, a method combining iterative nearest-point algorithm and principal component analysis is used to calculate the six-degree-of-freedom pose parameters of the workpiece in space. First, the workpiece body region is segmented from the 3D point cloud. The segmentation is based on a point cloud height threshold and connectivity analysis: background points lower than the support platform height are removed, and point clusters higher than the platform and connected to the main workpiece are retained. Next, outlier removal is performed on the segmented point cloud, with a neighborhood radius of 3 mm, retaining the set of internal points with a percentage greater than 95%. Then, the normal vector of each point is estimated using covariance matrix eigenvalue decomposition, with a neighborhood of 30 points. Using a standard CAD model point cloud as a reference template, coarse registration is performed using the iterative nearest-point algorithm. The iteration terminates when the mean square error is less than 0.2 mm or the number of iterations reaches fifty. Based on the coarse registration, a local optimization method based on principal component analysis is used: principal component analysis is performed on the point cloud of the workpiece's top surface region, extracting the first principal component direction as the local X-axis, the second principal component direction as the Y-axis, and the third principal component direction as the normal Z-axis. By comparing the angle between the actual point cloud's main axis and the CAD model's main axis, the slight warping caused by stamping springback is compensated. The final output is six-DOF pose parameters, including translations (x, y, z) along the world coordinate system's X, Y, and Z axes, and rotation angles (θ_x, θ_y, θ_z) around the X, Y, and Z axes, with pose calculation errors of less than 0.1 mm and 0.2 degrees, respectively.

[0028] The quality level determination result and the six-DOF pose parameters are input into the predictive command generation module. This module incorporates a kinematic model of the sorting actuator, including the complete time delay from receiving the command to the push rod being fully extended. . It consists of three parts: the servo valve electromagnetic response time (15 ms), the time required for the cylinder piston to accelerate to steady-state speed (12 ms), and the idle stroke time caused by mechanical transmission backlash (8 ms), for a total τ of 35 ms. The module acquires the workpiece's current position P_current on the conveyor belt in real time, calculated from the encoder's accumulated pulses, with an accuracy of ±0.02 mm. Combining the current speed v and pose parameters, it calculates the workpiece's position on the conveyor belt. Position after time Where v is the instantaneous linear velocity of the conveyor belt obtained in real time by the encoder, and P_current is the current position of the workpiece in the world coordinate system calculated based on the same encoder signal. Simultaneously, based on the predicted position and the workpiece posture, it is mapped to the spatial action point of the sorting actuator in the workpiece body coordinate system. The process of mapping the predicted position P_target and workpiece posture to the spatial action point in the workpiece body coordinate system includes: first, based on the calculated six-degree-of-freedom pose, the predicted position P_target in the world coordinate system is transformed to the workpiece body coordinate system through inverse rigid body transformation; second, in the workpiece body coordinate system, the optimal action point is automatically selected according to the predefined workpiece geometric model and sorting strategy. The selection rule for the optimal action point is: Located in the rigid support area of ​​the workpiece; Stay away from identified defect areas and workpiece edges; This maximizes the force arm of the push rod.

[0029] Final determination of target action time , and the coordinates of the point of action in space (x_act, y_act, z_act) and the extension direction vector d of the push rod.

[0030] The target action time and spatial point of action are directly transmitted to the servo controller of the sorting execution mechanism via a hard real-time communication bus. The hard real-time communication bus is an EtherCAT bus, with the master station cycle set at 500 microseconds and a time-triggered protocol to ensure that the command is transmitted within the specified cycle. The slave station consists of three independent servo drives, each controlling one of the three pneumatic push rod units. Each push rod unit is equipped with a proportional servo valve and a magnetostrictive displacement sensor with a displacement resolution of 0.01 mm. After receiving the command, the servo controller analyzes the target extension length l, the force F, and the trigger time t_action. The force F is dynamically calculated based on the workpiece mass, material thickness, and current posture: F = k × m × g / cos(α), where k is a safety factor of 1.2, m is the workpiece mass, g is the gravitational acceleration, and α is the angle between the push rod's action surface and the horizontal plane. The push rod end is equipped with a flexible silicone contact head with a Shore A60 hardness and a contact area of ​​20 square millimeters to ensure uniform force application on surfaces with different curvatures, avoiding indentation or slippage. At time t_action, the three push rods move sequentially according to a preset timing sequence with a timing difference of less than 2 milliseconds. The resultant force is precisely directed towards the center of the sorting channel, completing the dynamic sorting of qualified and defective products.

[0031] The system also includes a workpiece arrival confirmation module, consisting of a secondary photoelectric switch located 10 centimeters downstream of the imaging area. This switch detects whether a workpiece has passed within a predetermined time window. If no signal is detected, it is determined to be a missed detection or false trigger, and a compensation process is immediately initiated: the conveyor belt is paused, an auxiliary area scan camera is activated for static re-shooting, and defect identification and pose calculation are re-executed to ensure no workpiece is missed. All data processing units are deployed on an industrial embedded platform, with defect identification and pose calculation sharing the same graphics processor module. This module supports INT8 quantization inference, and the processing time for a single frame does not exceed 45 milliseconds, meeting the production line cycle requirement of 40 pieces per minute.

[0032] Each step of the above method is implemented through deep hardware and software coupling. The synchronous triggering unit is integrated into a field-programmable gate array chip, with solidified logic circuitry and deterministic response; the imaging unit uses an industrial-grade high-speed linear scan camera and a structured light projector to resist ambient light interference; the defect identification unit deploys a lightweight dual-stream network, balancing accuracy and speed; the pose calculation unit adopts a hybrid registration strategy, balancing global robustness and local accuracy; the predictive instruction generation module is based on measured dynamic parameters and non-idealized assumptions; the execution unit uses a multi-degree-of-freedom pneumatic array, possessing dual control capabilities for force and position; communication uses a hard real-time bus, eliminating the polling delay of traditional PLCs, and the entire system forms a closed loop of "perception-decision-execution".

[0033] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for quality sorting of automotive stamping parts combined with optical detection, characterized in that, include: An encoder and a photoelectric triggering device are set on the stamping part conveying path. The encoder is installed on the conveyor belt drive shaft and is used to output the conveyor belt running speed signal in real time. The photoelectric triggering device is arranged in the workpiece inlet area along the conveying direction to detect the leading edge position of the workpiece and generate an initial trigger signal. Based on the initial trigger signal and the speed signal output by the encoder, the time window for the workpiece to reach the preset imaging area is calculated, and a synchronous trigger command is sent to the high-speed linear scan camera and the structured light projector accordingly. This causes the high-speed linear scan camera to perform line-by-line exposure when the workpiece passes through the imaging area in a way that is relatively stationary with respect to the workpiece surface, while the structured light projector projects a multi-frequency phase-shifting stripe pattern onto the workpiece surface. The high-speed linear array camera acquires a sequence of structured light images reflected from the workpiece surface, and the three-dimensional point cloud data of the workpiece surface is calculated based on the phase unfolding algorithm. At the same time, a deep convolutional neural network is used to fuse the three-dimensional point cloud data with the corresponding two-dimensional grayscale image acquired by the high-speed linear array camera to extract the surface defect features of the workpiece and determine the quality level of the workpiece. Based on the three-dimensional point cloud data, the iterative nearest point algorithm combined with principal component analysis is used to calculate the six-degree-of-freedom pose parameters of the workpiece in space, including the translation along the X, Y, and Z axes and the rotation angles around the X, Y, and Z axes. The quality level determination result and the six-degree-of-freedom pose parameters are input into the predictive instruction generation module. This module calculates the target action time and spatial action point of the sorting actuator based on the real-time position of the current workpiece on the conveyor belt, the conveying speed, and the physical response delay of the downstream sorting actuator. The target action time and spatial point of action are directly sent to the servo controller of the sorting execution mechanism through a hard real-time communication bus. The servo controller drives the multi-degree-of-freedom pneumatic push rod to act on the specified position of the workpiece at a preset time with a specified force and direction, so as to complete the dynamic sorting of qualified products and defective products.

2. The method for quality sorting of automotive stamping parts combined with optical detection according to claim 1, characterized in that, The line frequency f of the high-speed linear array camera is determined by the following formula: ； in The conveyor belt speed is measured in real time by the encoder. The pixel physical size of the high-speed linear array camera is determined by the camera lens focal length, sensor pixel size, and object distance.

3. The method for quality sorting of automotive stamping parts combined with optical detection according to claim 2, characterized in that, The multi-frequency phase-shifted fringe pattern projected by the structured light projector includes three sets of sinusoidal fringes with different spatial frequencies. Each set contains four images with a phase difference of π / 2, which are used to calculate the wrapping phase using the four-step phase-shifting method and to perform phase unwrapping using the three-frequency heterodyne method to obtain the absolute phase map.

4. The method for quality sorting of automotive stamping parts combined with optical detection according to claim 3, characterized in that, The deep convolutional neural network is a two-stream encoder-decoder architecture, wherein the first stream input is a two-dimensional grayscale image, which extracts texture and edge features through convolutional layers and max pooling operations; the second stream input is three-dimensional point cloud data, which is voxelized and then extracted through three-dimensional convolutional layers to extract geometric features. The two-stream features are spliced ​​together at the bottleneck layer and then weighted and integrated through the attention fusion module to finally output a pixel-level defect classification map. The defect types include scratches, dents, wrinkles, oil stains and edge defects.

5. The method for quality sorting of automotive stamping parts combined with optical detection according to claim 4, characterized in that, The six-degree-of-freedom pose calculation process includes: firstly, segmenting the workpiece body region from the 3D point cloud and removing the background and supporting structure; denoising and normal vector estimation of the segmented point cloud; using the standard CAD model point cloud as a reference, coarse registration is performed through the iterative nearest point algorithm; based on the coarse registration, a local optimization method based on principal component analysis is used to compensate for the slight warping of the workpiece caused by stamping springback, and output accurate six-degree-of-freedom pose parameters.

6. The method for quality sorting of automotive stamping parts combined with optical detection according to claim 5, characterized in that, The predictive instruction generation module incorporates a kinematic model of the sorting actuator, which includes the complete time delay from receiving an instruction to completing the action. The predictive instruction generation module calculates the workpiece's position based on its current velocity v and pose. Predicted location after time The predicted position is combined with the workpiece posture and mapped to the spatial action point of the sorting execution mechanism in the workpiece body coordinate system, where P_current is the current position of the workpiece in the world coordinate system calculated based on the same encoder signal, and v is the instantaneous speed of the conveyor belt obtained in real time by the encoder.

7. A quality sorting system for automotive stamped parts incorporating optical detection, characterized in that, include: The synchronous triggering unit generates synchronous triggering commands for the camera and the structured light projector based on photoelectric triggering signals and encoder speed signals; The imaging unit includes a high-speed linear array camera and a multi-frequency structured light projector, used to acquire two-dimensional images and three-dimensional topographic data of the workpiece surface; The defect identification and quality assessment unit is used to detect and classify defects in the fused multimodal data through a deep convolutional neural network and output a quality level. A six-degree-of-freedom pose calculation unit calculates the complete pose parameters of a workpiece in space based on three-dimensional point cloud data; The predictive instruction generation module is used to calculate the target spatiotemporal parameters of the sorting action by combining pose parameters, transmission speed and actuator dynamics. The dynamic sorting execution unit includes a multi-degree-of-freedom pneumatic push rod array and its servo controller, which is used to perform sorting actions with specified parameters at a specified time; A hard real-time communication bus connects the predictive instruction generation module and the dynamic sorting execution unit, with a communication cycle of no more than 1 millisecond.

8. The automotive stamping parts quality sorting system combined with optical detection according to claim 7, characterized in that, The multi-degree-of-freedom push rod array includes three independently controlled pneumatic push rod units arranged along the conveying direction; each pneumatic push rod unit is equipped with a proportional servo valve and a displacement sensor, which can independently control the extension length, the force applied, and the triggering timing.

9. The automotive stamping parts quality sorting system combined with optical detection according to claim 7, characterized in that, The hard real-time communication bus is an EtherCAT bus, the master station cycle is set to 500 microseconds, and it supports distributed clock synchronization. The slave station is the servo controller.

10. The automotive stamping parts quality sorting system incorporating optical detection according to claim 9, characterized in that, The synchronous triggering unit is integrated into the industrial field programmable gate array chip. After receiving the initial trigger signal, its internal logic circuit immediately latches the current encoder count value and, based on the preset imaging distance and transmission speed, outputs a trigger pulse to the high-speed linear array camera and structured light projector at the calculated precise time through a hardware timer. The output trigger pulse jitter is less than 5 microseconds.

Images