A laser welding quality visual inspection system and method

Abstract

The present application relates to a kind of laser welding quality visual detection system and method, it relates to laser processing detection technical field, its system includes light source module, image acquisition module, image processing chip and control module group.The application is projected by light source module to welding area with the auxiliary illumination light of specific wavelength structured light, utilize image acquisition module real-time capture molten pool and the two-dimensional and three-dimensional feature image of weld, and send to image processing chip and carry out high-speed feature extraction and defect identification, finally by control module group the detection result is converted into control instruction and sent to laser welding execution mechanism, realize the on-line closed loop regulation of welding quality.The application solves the problem that traditional visual detection is susceptible to plasma arc light interference, lack of depth direction size detection, significantly improves the detection precision and real-time response speed of laser welding quality defect, guarantees the yield of welding manufacturing.

Description

Technical Field

[0001] This invention relates to the field of laser processing inspection technology, and in particular to a visual inspection system and method for laser welding quality. Background Technology

[0002] Laser welding, as an advanced manufacturing process characterized by high energy density, high precision, and high efficiency, has been widely applied in fields such as automotive manufacturing, aerospace, new energy power batteries, and precision electronic machining. During laser welding, due to the combined effects of complex factors such as uneven material surface conditions, fluctuations in workpiece assembly gaps, laser output power fluctuations, and environmental vibrations, various welding defects, including porosity, spatter, incomplete penetration, undercut, and burn-through, are easily generated at the weld joint. The presence of these welding defects directly impairs the mechanical strength, fatigue resistance, and sealing performance of the weld joint, and can even pose serious threats to the safety and operational reliability of the entire equipment. Therefore, high-precision real-time quality inspection during the welding process is crucial.

[0003] Currently, to ensure the final quality of laser welding, the industry typically employs offline inspection or online monitoring methods. While offline inspection methods such as ultrasonic testing, X-ray inspection, and metallographic section analysis offer high accuracy, they suffer from long inspection cycles, high costs, and the inability to provide real-time online quality feedback, making them unsuitable for real-time adjustments in large-scale production lines. Existing online visual inspection technologies largely rely on single-channel 2D CCD or CMOS industrial cameras coupled with conventional backlights to perform 2D image analysis of the weld pool area. However, in actual laser welding processes, high-intensity laser radiation, glaring metallic plasma arc light, and spatter generate significant background noise, making images captured by single-channel 2D cameras prone to localized overexposure or loss of detail.

[0004] Because the image sensors in the existing technologies are easily affected by the strong light of the molten pool and the plasma arc light, and traditional two-dimensional visual images lack effective characterization of three-dimensional geometric features such as molten pool depth and weld reinforcement, the existing visual inspection systems suffer from poor detection accuracy, high false alarm rate and false alarm rate, and the inability to judge welding quality from the depth dimension. These major technical problems greatly limit the quality control level of automated laser welding production. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a visual inspection system and method for laser welding quality.

[0006] This application provides a laser welding quality visual inspection system, which adopts the following technical solution: A laser welding quality visual inspection system includes: a light source module, an image acquisition module, an image processing chip, and a control module connected to the image processing chip; the light source module is used to emit structured light and auxiliary illumination light to the laser welding area; the image acquisition module is used to acquire two-dimensional images and three-dimensional morphological features of the laser welding molten pool and weld in real time, and transmit them to the image processing chip; the image processing chip is used to preprocess the received images, extract features, and identify quality defects to obtain inspection results; the control module outputs control commands to the laser welding actuator based on the inspection results.

[0007] By adopting the above technical solution, the light source module provides a multi-source illumination environment with specific incident angles and wavelengths. Even under strong arc light interference, the image acquisition module can still acquire high-quality two-dimensional textures and three-dimensional structures of the molten pool. The image processing chip performs parallel calculations, filtering, and classification on the acquired image data, and the control module implements closed-loop feedback control of the actuator. These modules work together to effectively solve the technical problems of traditional visual inspection being susceptible to strong light interference and lacking three-dimensional features, achieving efficient and high-precision online inspection of laser welding quality.

[0008] Preferably, the light source module includes a coaxial ring light source and an oblique multi-band structure light source. The coaxial ring light source is configured as a high color rendering LED light source, and the oblique multi-band structure light source is configured as a blue semiconductor laser generator.

[0009] By employing the above technical solution, the coaxial ring light source provides shadowless, uniform illumination to highlight the edge features of the weld in the two-dimensional direction, while the oblique-projection multi-band structured light source projects multi-line or single-line structured light stripes to support three-dimensional morphology reconstruction. The combination of these two technologies and their complementary wavelength design can strongly suppress the plasma arc light in the molten pool, greatly improving the signal-to-noise ratio of the image.

[0010] Preferably, the image acquisition module includes a high frame rate industrial camera and a narrowband filter, wherein the center cutoff wavelength of the narrowband filter matches the wavelength of the oblique multi-band structure light source, and the narrowband filter is configured as an 850nm infrared narrowband filter.

[0011] By adopting the above technical solution, the high frame rate industrial camera can capture the high-speed dynamic changes of the molten pool, while the narrow band filter only allows structured light and auxiliary light of specific wavelengths to pass through, filtering out most of the stray background light such as laser radiation and metal arc light, thus ensuring high contrast and clarity of the molten pool image.

[0012] Preferably, the image processing chip includes an FPGA preprocessing unit and a DSP core processing unit. The FPGA preprocessing unit is used to perform image denoising, Gaussian filtering, and coarse edge extraction, while the DSP core processing unit is used to perform molten pool 3D reconstruction, defect classification, and size measurement.

[0013] By adopting the above technical solution, the FPGA preprocessing unit leverages its hardware parallel pipeline processing advantages to perform high-speed denoising and coarse extraction on massive image data, while the DSP core processing unit is responsible for complex floating-point operations, 3D height calculations, and advanced classification, thus constructing an efficient architecture of "high-speed preprocessing in the front stage + precise calculation in the back stage" to ensure the real-time performance of the system detection.

[0014] Preferably, the control module includes an Ethernet communication unit and a PLC execution controller, wherein the Ethernet communication unit uses the EtherCAT protocol to perform real-time data interaction with the laser welding actuator.

[0015] By adopting the above technical solution, the high-speed bus protocol of the Ethernet communication unit ensures extremely low latency transmission of detection results and control commands, while the PLC execution controller can quickly and reliably adjust the laser power or welding speed, realizing high-precision and high-response physical closed-loop control.

[0016] This application also provides a detection method for a laser welding quality visual inspection system.

[0017] A detection method for a laser welding quality visual inspection system includes: Step S1: Obtain a real-time image sequence of the laser welding molten pool through the image acquisition module; Step S2: Filter and denoise the image sequence and segment the ROI region to extract the molten pool boundary and weld centerline; Step S3: Calculate the three-dimensional height map of the molten pool surface based on the optical triangulation principle; Step S4: Use deep learning classification algorithms to identify defects in the two-dimensional features and three-dimensional height map of the molten pool, and obtain the welding defect type and location; Step S5: Based on the defect type and location, control the laser welding actuator to perform parameter compensation or shutdown alarm.

[0018] By adopting the above technical solution, image acquisition, preprocessing, 3D height calculation, deep learning intelligent classification, and closed-loop feedback control are integrated into a complete automated detection process. Based on the principle of optical triangulation, 3D features are introduced, compensating for the lack of dimension in traditional 2D vision. Combined with deep learning algorithms, this significantly improves the accuracy and efficiency of identifying complex welding defects.

[0019] Preferably, in step S2, the filtering and denoising adopts a dynamic adaptive median filtering algorithm, and the ROI region segmentation adopts an adaptive threshold segmentation based on the maximum inter-class variance method.

[0020] By adopting the above technical solutions, dynamic adaptive median filtering or bilateral filtering can effectively filter out high-frequency pulse noise and electromagnetic interference while well preserving the edge abrupt change information of the molten pool and weld; Otsu's adaptive threshold segmentation can automatically set the segmentation threshold according to the local gray-scale distribution of the image, ensuring the accurate segmentation of the region of interest (ROI).

[0021] Preferably, the calculation of the three-dimensional height map of the molten pool surface in step S3 includes: obtaining the stripe center line using a feature point matching algorithm, and calculating the height difference based on the calibration parameter matrix, wherein the calibration parameter matrix is ​​obtained by the Zhang Zhengyou calibration method.

[0022] By adopting the above technical solution, through the extraction of the stripe center line and the conversion of the high-precision calibration matrix, the pixel coordinate displacement on the image can be accurately converted into the height fluctuations in the physical world, realizing the sub-millimeter level measurement of the three-dimensional depth of the molten pool and the surface residual height.

[0023] Preferably, the deep learning classification algorithm in step S4 is the MobileNetV3 network with an attention mechanism.

[0024] By adopting the above technical solutions, the lightweight MobileNetV3 network significantly reduces the number of model parameters and computational overhead while maintaining high recognition accuracy. Combined with the attention mechanism, it can automatically focus on the key defect feature areas of the melt pool, balancing detection accuracy with the real-time operation requirements of embedded devices.

[0025] Preferably, the parameter compensation in step S5 includes: calculating the laser power compensation value and the welding speed compensation value based on the weld pool width deviation, wherein the compensation value is calculated by an incremental PID algorithm.

[0026] By adopting the above technical solution, the incremental PID algorithm is used to dynamically adjust the geometric deviation of the molten pool in a closed loop. This allows for the rapid calculation of the optimal laser power and speed adjustment commands, eliminating transient disturbances during the welding process and ensuring the consistency of the final weld formation.

[0027] In summary, this application includes at least one of the following beneficial technical effects: 1. The light source module provides a multi-source illumination environment with specific incident angles and wavelengths. Even under strong arc light interference, the image acquisition module can still acquire high-quality 2D textures and 3D structures of the molten pool. The image processing chip performs parallel calculations, filtering, and classification on the acquired image data. The control module implements closed-loop feedback control of the actuators. These modules work together to effectively solve the technical problems of traditional visual inspection being susceptible to strong light interference and lacking 3D features, achieving efficient and high-precision online inspection of laser welding quality.

[0028] 2. Image acquisition, preprocessing, 3D height calculation, deep learning intelligent classification, and closed-loop feedback control are integrated into a complete automated detection process. Based on the principle of optical triangulation, 3D features are introduced to compensate for the lack of dimension in traditional 2D vision. Combined with deep learning algorithms, this significantly improves the accuracy and efficiency of identifying complex welding defects. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall structure of a laser welding quality visual inspection system.

[0030] Figure 2 This is a schematic diagram of the geometric triangulation measurement principle between the light source module and the image acquisition module.

[0031] Figure 3 This is a block diagram showing the internal architecture and interconnections of an image processing chip.

[0032] Figure 4 This is a flowchart illustrating a visual inspection method for laser welding quality.

[0033] Explanation of reference numerals in the attached diagram: 1. Light source module; 11. Ring light source; 12. Oblique-type multi-band structure light source; 2. Image acquisition module; 3. Image processing chip; 31. FPGA preprocessing unit; 32. DSP core processing unit; 4. Control module. Detailed Implementation

[0034] The present application will be further described in detail below with reference to all the accompanying drawings.

[0035] This application discloses a laser welding quality visual inspection system.

[0036] Reference Figures 1 to 3 A laser welding quality visual inspection system includes: a light source module 1, an image acquisition module 2, an image processing chip 3, and a control module 4 connected to the image processing chip 3.

[0037] In the specific spatial structure and physical connection, the light source module 1 is rigidly installed on the lower outer side of the laser welding head, and its projection axis forms a preset angle with the optical axis of the laser beam. The light source module 1 includes a coaxial ring light source 11 and an oblique multi-band structure light source 12. In this embodiment, the coaxial ring light source 11 is configured as a high color rendering LED light source (for example, a white LED ring with a color rendering index greater than 90, or a high-brightness monochromatic red LED ring). It is installed below the laser collimator and coaxial with the laser beam, emitting uniformly distributed auxiliary illumination light vertically downwards to uniformly illuminate the weld and the base material around the molten pool in the laser welding area, thereby eliminating shadows and facilitating the extraction of weld edges. The oblique-projection multi-band structured light source 12 is installed on the front or rear side of the welding direction, and the angle between its optical axis and the horizontal plane is configured to be between 30 degrees and 60 degrees (e.g., 45-degree oblique projection). The oblique-projection multi-band structured light source 12 is configured as a blue semiconductor laser generator (e.g., a blue line structured light laser with a wavelength of 450nm and a power of 100mW) or a green semiconductor laser generator (e.g., a green line structured light laser with a wavelength of 532nm and a power of 80mW) to project high-brightness linear or grid-like structured light stripes onto the surface of the molten pool. The structured light stripes bend and deform with the undulations of the molten pool surface, thereby carrying three-dimensional height information.

[0038] Image acquisition module 2 is rigidly fixed to the other side of the laser welding head, and its optical axis is also set at a certain angle with the laser beam optical axis to form a triangulation geometry. Image acquisition module 2 includes a high frame rate industrial camera and a narrowband filter. The high frame rate industrial camera is configured as a high-speed CMOS camera with a gigabit network interface or a CameraLink interface (e.g., a frame rate of not less than 500 frames / second at full resolution, and a resolution of 1280×1024 pixels), used to capture image sequences of dynamic changes in the laser welding molten pool in real time. The narrowband filter is threadedly connected to the front end of the lens of the high-speed industrial camera. The center cutoff wavelength of the narrowband filter matches the wavelength of the oblique multi-band structured light source 12. The narrowband filter is configured as an 850nm infrared narrowband filter (when the structured light source is infrared or utilizes infrared molten pool self-radiation) or a 450nm blue light narrowband filter (when the structured light source is a 450nm blue laser). The bandwidth of the narrowband filter can be selected as 10nm or 20nm. Because laser welding generates intense metal vapor plasma arc light (mainly concentrated in the ultraviolet to visible light band) and high-energy reflected laser light (such as the commonly used fiber laser with a wavelength of 1064nm), the matching narrowband filter can filter out most of the stray strong light except for the structured light wavelength, so that the camera's image sensor mainly presents a clear contrast pattern between the structured light stripes and the molten pool.

[0039] Image processing chip 3 is connected to the industrial camera of image acquisition module 2 via a high-speed data bus. Image processing chip 3 adopts a heterogeneous computing architecture, including an FPGA preprocessing unit 31 and a DSP core processing unit 32. The FPGA preprocessing unit 31 (e.g., using a Xilinx Kintex-7 series chip) performs real-time noise reduction and Gaussian filtering on the raw pixel stream from the camera in parallel through logic circuit hardware, and performs primary edge coarse extraction and region of interest (ROI) cropping to significantly reduce the amount of data for subsequent processing. The DSP core processing unit 32 (e.g., using a TI TMS320C6678 multi-core digital signal processor) is interconnected with the FPGA via a high-speed SRIO (Serial Rapid I / O) bus, and is used to receive the feature data processed by the FPGA, and to perform high-precision molten pool surface 3D reconstruction algorithms, defect classification deep learning inference, and precise measurement of molten pool geometric feature dimensions.

[0040] The input terminal of control module 4 is connected to the output interface of image processing chip 3. Control module 4 includes an Ethernet communication unit and a PLC execution controller. The Ethernet communication unit adopts a hard-core Ethernet controller, supports EtherCAT or Profinet protocols, and establishes a real-time data interaction channel with external laser welding actuators (such as six-axis industrial robot controllers or 3D CNC systems). The PLC execution controller (e.g., using a Siemens S7-1500 series PLC) sends control commands in real time to the laser controller and the motion mechanism driver via the bus based on the defect type and size deviation results sent by image processing chip 3, dynamically adjusting the laser power, welding speed, or shielding gas flow rate. The working principle of a laser welding quality visual inspection system according to an embodiment of this application is as follows: When the laser welding system is started, the laser beam is focused and irradiates the surface of the workpiece to be welded, forming a high-temperature molten pool. A coaxial ring light source 11 and an oblique multi-band structured light source 12 are simultaneously turned on, projecting auxiliary illumination light and linear structured light stripes onto the welding molten pool and surrounding area. A high-frame-rate industrial camera, protected by a narrow-band filter, filters out intense plasma arc light and continuously captures image sequences of the structured light stripes deforming on the surface of the molten pool. These images are then fed in real-time into the FPGA preprocessing unit 31 for high-speed Gaussian denoising and ROI segmentation. Subsequently, the preprocessed image is sent to the DSP core processing unit 32. The DSP uses the principle of optical triangulation geometry to calculate the three-dimensional height of each point on the surface of the molten pool based on the bending displacement of the stripes, reconstructing the three-dimensional morphological features of the molten pool and weld, and extracting feature vectors such as the width, height, area, and symmetry of the molten pool. Next, the DSP runs a deep learning classification model to infer the presence of defects such as porosity, burn-through, or incomplete penetration from the feature map combining two-dimensional texture and three-dimensional height. If a defect trend is detected, control module 4 immediately sends compensation control commands to the robot controller and laser via the EtherCAT industrial bus to fine-tune the laser power and speed in real time, thereby achieving online closed-loop control of welding quality and ensuring the stability of the welding process.

[0041] The application also discloses a detection method for a laser welding quality visual inspection system.

[0042] Reference Figure 4 A detection method for a laser welding quality visual inspection system includes the following steps: S1: Acquire real-time image sequences of the laser welding molten pool through the image acquisition module.

[0043] In practice, the industrial camera captures images of the molten pool, the front seam to be welded, and the rear solidified seam simultaneously as the laser welding actuator moves the welding head along the weld seam at a data acquisition rate of 1000 frames per second.

[0044] S2: Filter and denoise the image sequence and segment the ROI region to extract the molten pool boundary and weld centerline.

[0045] Due to strong electromagnetic interference and residual arc flicker at the welding site, the acquired images contain high-frequency pulse noise. In step S2, dynamic adaptive median filtering or bilateral filtering algorithms are used for denoising.

[0046] In this embodiment, if a dynamic adaptive median filtering algorithm is used, its basic principle is to dynamically adjust the size of the filtering window based on the pixel extreme values ​​within the local window. The calculation process is as follows: First, define the maximum value of pixels within the window. Minimum value and median .like If the median is not noise, then proceed to determine the current center pixel. Is it noise? like Then output Otherwise output If not satisfied Increase the window size. Continue judging until the window reaches the set maximum value (e.g., 7×7), then output directly. .

[0047] If a bilateral filtering algorithm is used, it can effectively preserve edges while filtering out noise, and its filtered output pixel values... The calculation formula is: ; in, The input pixel value is the weight factor within the neighborhood. It is a spatial neighborhood factor Similarity factor with pixel value The product of: ; In the formula, This is the standard deviation in the spatial domain, used to control the geometric range of the filter (for example, it can be set to 3.0). This is the standard deviation of the range, used to control the sensitivity to edge strength (for example, it can be set to 30.0).

[0048] After filtering, to reduce computational overhead in unnecessary areas, an adaptive threshold segmentation algorithm based on the Otsu method (maximum inter-class variance method) is used to locate the melt pool region and crop out the region of interest (ROI) containing the melt pool and structured light fringes. The Otsu method automatically determines the segmentation threshold by calculating the maximum inter-class variance between the background and foreground. : ; In the formula, and Thresholds The proportion of background and foreground pixels in the entire image. and These are the average grayscale values ​​for the background and foreground, respectively. This threshold is then used to determine the grayscale values. After binarizing the image, the molten pool boundary can be accurately segmented, and the weld centerline can be extracted using a skeleton extraction algorithm.

[0049] S3: Calculate the three-dimensional height map of the molten pool surface based on the optical triangulation principle.

[0050] Step S3, which calculates the three-dimensional height map of the molten pool surface, includes: obtaining the stripe center line using a feature point matching algorithm, and calculating the height difference based on the calibration parameter matrix.

[0051] The principle of optical triangulation is as follows: Let the angle between the structured light projector and the camera's optical axis be... The angle between the projector's optical axis and the reference plane of the object being measured is When the surface of the molten pool changes significantly due to welding defects or molten pool fluctuations. At that time, the pixel displacement generated on the camera's image sensor by the structured light stripes projected onto the molten pool surface is (Unit: pixels). Combined with camera magnification. (Unit: mm / pixel), physical displacement is .

[0052] Calculate the height variation of points on the molten pool surface based on trigonometric relationships. The calculation formula is: ; In the formula, This indicates the height difference (unit: mm) between the measured point on the surface of the molten pool and the reference horizontal plane. This indicates the pixel deviation of the point in the image relative to the invariant reference fringe; The horizontal resolution scaling factor of the camera; This indicates the angle between the camera and the structured light source (e.g., it can be set to 45°). This indicates the incident angle of the structured light (e.g., it can be set to 45°).

[0053] To obtain high-precision calculation results, the system must perform camera calibration beforehand. The calibration parameter matrix is ​​obtained through Zhang Zhengyou's calibration method or binocular vision calibration. The calibrated camera intrinsic parameter matrix... extrinsic parameter matrix In addition, the lens distortion coefficient can accurately transform the image pixel coordinates to the world coordinate system, thereby eliminating the influence of the radial and tangential distortion of the lens on the height calculation and realizing the reconstruction of the 3D height map of the molten pool at the sub-millimeter level (such as ±0.05mm).

[0054] S4: Use deep learning classification algorithms to identify defects in the two-dimensional features and three-dimensional height map of the molten pool.

[0055] Determine if a defect exists. If a defect exists, obtain the type and location of the welding defect and proceed to step S5.

[0056] If no defects are found, repeat steps S1 through S4.

[0057] In step S4, the deep learning classification algorithm is either a MobileNetV3 network or a ResNet18 network that incorporates an attention mechanism.

[0058] In this embodiment, the input data is a multi-channel feature map that fuses a 2D grayscale image of the melt pool and a 3D height map. If a lightweight MobileNetV3 network with a fusion attention mechanism (such as CBAM, convolutional block attention module) is used, its network structure mainly consists of depthwise separable convolution and inverted residuals, and embeds channel attention mechanism and spatial attention mechanism within it.

[0059] The channel attention mechanism compresses the spatial dimension through global average pooling and max pooling. The calculation formula is as follows: ; In the formula, Given the input feature map, It is a multilayer perceptron. It is the Sigmoid activation function. The channel attention weights for the output.

[0060] The spatial attention mechanism performs average pooling and max pooling along the channel dimension, and then performs convolution operations: ; In the formula, This represents a convolution kernel of size 7×7.

[0061] Through the CBAM module, the network can autonomously learn and focus on areas with strong distortion at the edge of the molten pool (corresponding to defects such as undercut and incomplete penetration) or areas with highly abrupt changes (corresponding to defects such as spatter and burn-through). While maintaining the network's extremely high computing speed (inference time less than 5ms), the accuracy of defect identification is improved to over 99.2%.

[0062] S5: Based on the type and location of the defect, control the laser welding actuator to perform parameter compensation or shutdown alarm.

[0063] The parameter compensation in step S5 includes: calculating the laser power compensation value and the welding speed compensation value based on the deviation of the molten pool width. The compensation value is calculated by an incremental PID algorithm or a fuzzy control algorithm.

[0064] In this embodiment, if an incremental PID control algorithm is used, its calculation formula is: ; In the formula, For the first The control output increment at each sampling time (such as the laser power adjustment increment) Or adjust the welding speed increment ; The current value is the deviation of the molten pool width (the difference between the standard width and the actual detection width). and These are the deviation values ​​for the previous two times, respectively; This is a scaling factor (e.g., set to 0.8). This is the integral coefficient (e.g., set to 0.05). This is the differential coefficient (e.g., set to 0.1).

[0065] Obtained through calculation The PLC controller immediately sends a power adjustment command to the laser or a speed fine-tuning command to the robot controller. For example, when it detects that the weld pool width is narrowing or the height is decreasing (indicating insufficient heat input and a tendency for incomplete penetration), the system automatically increases the laser power. Or reduce welding speed If severe sinking of the molten pool (burn-through trend) is detected, a shutdown alarm will be triggered immediately and the laser will be shut down, thereby achieving efficient closed-loop protection of welding quality.

[0066] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A laser welding quality visual inspection system, characterized in that, include: The light source module (1), the image acquisition module (2), the image processing chip (3), and the control module (4) connected to the image processing chip (3); The light source module (1) is used to emit structured light and auxiliary illumination light to the laser welding area; The image acquisition module (2) is used to acquire two-dimensional images and three-dimensional morphological features of the laser welding molten pool and weld in real time, and transmit them to the image processing chip (3); The image processing chip (3) is used to preprocess the received image, extract features and identify quality defects to obtain the detection result; The control module (4) outputs control commands to the laser welding actuator based on the detection results.

2. The laser welding quality visual inspection system according to claim 1, characterized in that, The light source module (1) includes a coaxial ring light source (11) and an oblique multi-band structure light source (12). The coaxial ring light source (11) is configured as a high color rendering LED light source, and the oblique multi-band structure light source (12) is configured as a blue semiconductor laser generator.

3. The laser welding quality visual inspection system according to claim 1, characterized in that, The image acquisition module (2) includes a high frame rate industrial camera and a narrowband filter. The center cutoff wavelength of the narrowband filter is matched with the wavelength of the oblique multi-band structure light source (12). The narrowband filter is configured as an 850nm infrared narrowband filter.

4. The laser welding quality visual inspection system according to claim 1, characterized in that, The image processing chip (3) includes an FPGA preprocessing unit (31) and a DSP core processing unit (32). The FPGA preprocessing unit (31) is used to perform image denoising, Gaussian filtering and edge coarse extraction. The DSP core processing unit (32) is used to perform molten pool three-dimensional reconstruction, defect classification and size measurement.

5. The laser welding quality visual inspection system according to claim 1, characterized in that, The control module (4) includes an Ethernet communication unit and a PLC execution controller. The Ethernet communication unit uses the EtherCAT protocol to perform real-time data interaction with the laser welding actuator.

6. A detection method based on the laser welding quality visual inspection system according to any one of claims 1 to 5, characterized in that, include: Step S1: Obtain a real-time image sequence of the laser welding molten pool through the image acquisition module (2); Step S2: Filter and denoise the image sequence and segment the ROI region to extract the molten pool boundary and weld centerline; Step S3: Calculate the three-dimensional height map of the molten pool surface based on the optical triangulation principle; Step S4: Use deep learning classification algorithms to identify defects in the two-dimensional features and three-dimensional height map of the molten pool, and obtain the welding defect type and location; Step S5: Based on the defect type and location, control the laser welding actuator to perform parameter compensation or shutdown alarm.

7. The laser welding quality visual inspection method according to claim 6, characterized in that, In step S2, the filtering and denoising adopts a dynamic adaptive median filtering algorithm, and the ROI region segmentation adopts an adaptive threshold segmentation based on the maximum inter-class variance method.

8. The laser welding quality visual inspection method according to claim 6, characterized in that, The calculation of the three-dimensional height map of the molten pool surface in step S3 includes: obtaining the stripe center line using a feature point matching algorithm, and calculating the height difference based on the calibration parameter matrix, which is obtained by the Zhang Zhengyou calibration method.

9. A visual inspection method for laser welding quality according to claim 6, characterized in that, In step S4, the deep learning classification algorithm is the MobileNetV3 network with an attention mechanism.

10. A visual inspection method for laser welding quality according to claim 6, characterized in that, The parameter compensation in step S5 includes: calculating the laser power compensation value and the welding speed compensation value based on the deviation of the weld pool width. The compensation value is calculated by an incremental PID algorithm.

Images