A laser welding target recognition and intelligent closed-loop control system and method

Abstract

The application discloses a laser welding target recognition and intelligent closed-loop control system and method, relates to the technical field of laser welding automatic control, and comprises a coaxial laser welding visual perception unit, an intelligent image enhancement and CNN target recognition unit, an intelligent closed-loop control unit and an execution unit which are sequentially connected in signal, is equipped with a specific wavelength oblique illumination light source, adopts a two-stage image enhancement architecture of MSRCR and light-weight DeblurGAN-v2 cascade, outputs pixel-level deviation data through an improved light-weight CNN, realizes self-adaptive PID control through a multi-input CNN-BiLSTM network optimized by Bayes, and drives a welding laser head to complete weld tracking, so that the recognition precision and real-time performance under strong interference can be improved, and the control stability under multiple working conditions can be enhanced.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for laser welding, and in particular to a laser welding target recognition and intelligent closed-loop control system and method. Background Technology

[0002] With the rapid development of industries such as high-end equipment manufacturing, new energy power batteries, and precision electronic devices, laser welding has become a core process for precision component welding due to its technological advantages, including high energy density, small heat-affected zone, high welding precision, and ease of automation integration. The increasingly demanding requirements of high-speed and precision laser welding production scenarios place increasingly stringent demands on the accuracy of automatic weld tracking and the response speed of closed-loop control: the strong arc light, plasma, and spatter generated during welding can significantly interfere with visual imaging, greatly increasing the difficulty of target feature extraction; simultaneously, the welding conditions fluctuate dynamically, and changes in parameters such as welding material, welding speed, and laser power can all affect welding quality, requiring the control system to have rapid response and adaptive adjustment capabilities to ensure consistent welding quality.

[0003] Currently, laser welding seam tracking technology is mainly based on visual sensing solutions, and its technological evolution can be divided into two categories: traditional machine vision solutions and deep learning-assisted solutions. Traditional machine vision solutions mostly use structured light sensors combined with edge detection algorithms to extract weld seam features. In stable welding scenarios, they can achieve basic weld seam tracking functions. However, when faced with complex conditions such as strong arc light interference, highly reflective base materials, and small-gap weld seams, the accuracy of feature extraction drops significantly, making it difficult to meet the processing requirements of precision welding. To improve the robustness of recognition under complex conditions, some existing technologies introduce convolutional neural networks into the welding target recognition stage. Deep learning models are used to extract features from the weld seam and laser spot, which improves the stability of recognition under complex conditions to some extent.

[0004] At the image enhancement preprocessing level, existing solutions mostly use a single algorithm to complete the illumination enhancement processing. Although this can improve the overall brightness of the image, it will introduce obvious edge halos and background noise around strong light areas such as laser spots, which will destroy the gradient characteristics of the weld edge and restrict the further improvement of subsequent recognition and positioning accuracy. The effect is difficult to meet the application requirements of high-precision recognition.

[0005] At the target recognition and localization level, existing deep learning recognition models mostly directly apply general detection network architectures, resulting in a large number of model parameters and high inference latency, making it difficult to adapt to high-speed welding application scenarios. If the model is overly lightweighted, it will lead to insufficient positioning accuracy, failing to meet the sub-millimeter positioning requirements of precision welding, and making it difficult to achieve an effective balance between detection speed and positioning accuracy.

[0006] In addition, existing technical solutions mostly focus on single-point improvements in the identification or control links, without building a complete intelligent closed-loop control link. The information interaction between the sensing and control links is insufficient, and the overall system lacks full-link coordination, making it difficult to achieve full-process collaborative optimization in high-speed precision welding scenarios.

[0007] In summary, existing laser welding target recognition and intelligent closed-loop control technologies still have significant shortcomings in terms of anti-interference capability under complex working conditions, balance between algorithm accuracy and real-time performance, multi-condition adaptive control, and end-to-end closed-loop collaboration, making it difficult to meet the production application needs of high-end precision laser welding. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a laser welding target recognition and intelligent closed-loop control system and method.

[0009] This invention is achieved through the following technical solution: A laser welding target recognition and intelligent closed-loop control system includes a specific wavelength oblique illumination source, a coaxial laser welding visual perception unit, an intelligent image enhancement and CNN target recognition unit, an intelligent closed-loop control unit, and an execution unit; each unit is sequentially connected to form a complete intelligent closed-loop control link. The specific wavelength oblique illumination source is made of a single-wavelength LED that is different from the wavelength of the welding laser, and is used to provide auxiliary illumination to the weld area; The coaxial laser welding vision perception unit is coaxially mounted on the reserved camera interface of the welding laser head, and includes an industrial lens, a narrowband filter, and a vision sensor arranged in sequence. The light transmission wavelength of the narrowband filter matches the emission wavelength of the specific wavelength oblique illumination source, and is used to filter out the strong arc light and plasma interference generated by the welding laser. The vision sensor is used to acquire the laser spot and weld target image during the laser welding process, and transmit the target image to the intelligent image enhancement and CNN target recognition unit. The intelligent image enhancement and CNN target recognition unit is used to perform intelligent enhancement preprocessing and pixel-level target recognition and positioning on the acquired image in sequence, and output the relative position deviation data between the laser spot and the weld seam. The intelligent closed-loop control unit incorporates a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network, which combines the relative position deviation data from visual feedback with real-time welding condition information to perform fuzzy closed-loop control calculations, outputs control signals to the execution unit, and iteratively optimizes control parameters based on real-time welding data. The execution unit includes a perpendicular X-axis motion component and a Y-axis motion component. Both the X-axis motion component and the Y-axis motion component include a servo motor and a lead screw slide. The execution unit is used to adjust the spatial position of the welding laser head in real time according to the control signal so that the laser beam is always aligned with the center of the weld.

[0010] According to the above technical solution, preferably, the intelligent image enhancement and CNN target recognition unit includes a laser welding image intelligent enhancement module and an improved lightweight CNN recognition module that are connected in sequence. The laser welding image intelligent enhancement module is used to perform defogging, noise reduction and feature enhancement processing on the input image to highlight the laser spot and weld edge features; The improved lightweight CNN recognition module is used to perform pixel-level recognition and positioning of weld edges and laser spots in the enhanced image. Its recognition targets include the center of the laser spot, the spot area and the edge of the weld, and outputs the relative position deviation data of the center of the laser spot relative to the weld.

[0011] According to the above technical solution, preferably, the laser welding image intelligent enhancement module includes an MSRCR enhancement submodule and a lightweight DeblurGAN-v2 correction submodule connected in sequence. The MSRCR enhancement submodule employs a multi-scale Retinex algorithm with color restoration to enhance the illumination of the input image and eliminate image degradation caused by smoke and spatter during welding. The lightweight DeblurGAN-v2 correction submodule is used to eliminate image noise introduced after enhancement processing by the MSRCR algorithm, as well as edge halos in the strong light area around the laser spot, and output the final enhanced image.

[0012] According to the above technical solution, preferably, the generative adversarial network of the lightweight DeblurGAN-v2 correction submodule adopts the WGAN-GP loss function with gradient penalty, its backbone network is replaced with MobileNet-DSC, and it has been pre-processed by model pruning and output quantization to reduce the amount of computation, shorten the processing time, and meet the real-time requirements of the laser welding process.

[0013] According to the above technical solution, preferably, the lightweight DeblurGAN-v2 correction submodule uses the laser welding image containing halo and noise after being processed by the MSRCR algorithm as a training sample to directionally repair the halo and background noise defects at the edge of the spot introduced by the MSRCR algorithm enhancement processing, while preserving the gradient features of the weld edge.

[0014] According to the above technical solution, preferably, the intelligent closed-loop control unit includes an adaptive PID controller; the input of the Bayesian optimized multi-input CNN-BiLSTM convolutional neural network includes relative position deviation data from visual feedback, welding material parameters, welding speed parameters, and laser power parameters, and the output is PID control parameters, which are used to achieve adaptive adjustment of control parameters under different welding conditions.

[0015] This invention also discloses a laser welding target recognition and intelligent closed-loop control method, which employs the aforementioned laser welding target recognition and intelligent closed-loop control system and includes the following steps executed cyclically: S1. A specific wavelength oblique illumination source is used to illuminate the weld area of ​​the workpiece at an oblique angle. The image is acquired by a coaxially mounted laser welding vision perception unit and narrowband optical filtering is performed to obtain a laser spot and weld target image that suppresses the strong arc light and plasma interference of the welding laser. S2. Based on the laser spot and weld target image obtained in step S1, the laser spot and weld target image are sequentially processed by intelligent enhancement preprocessing and improved lightweight convolutional neural network recognition and localization processing to obtain pixel-level relative position deviation data between the laser spot and the weld. S3. Based on the pixel-level relative position deviation data obtained in step S2, combined with the real-time welding condition information of welding material, welding speed, and laser power, adaptive fuzzy closed-loop control calculation is performed through a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network to generate corresponding control signals and output them to the execution unit to drive the execution unit to adjust the spatial position of the welding laser head. S4. Real-time acquisition of relative position deviation data and execution unit action signals during the welding process and status monitoring and judgment; when the relative position deviation data exceeds the set threshold or the execution unit malfunctions, an alarm is triggered and the welding operation is suspended; under normal working conditions, the real-time operating data is fed back to the acquisition and control link, and the execution step S1 is returned to form a continuous dynamic closed-loop control.

[0016] According to the above technical solution, preferably, the intelligent enhancement preprocessing in step S2 specifically includes the following sub-steps: S21. Based on the input target image, the illumination enhancement process is performed using the multi-scale Retinex algorithm with color restoration to obtain a preliminary enhanced image that eliminates the degrading effects of smoke and splashes. S22. Based on the preliminary enhanced image obtained in step S21, noise and halo removal processing is performed using a lightweight DeblurGAN-v2 generative adversarial network to obtain the final enhanced image.

[0017] According to the above technical solution, preferably, the lightweight DeblurGAN-v2 generative adversarial network in step S22 adopts the WGAN-GP loss function with gradient penalty, its backbone network is replaced with MobileNet-DSC, and it has been pre-processed with model pruning and output quantization to reduce computational latency and meet the real-time requirements of the laser welding process.

[0018] According to the above technical solution, preferably, in step S3, the Bayesian optimized multi-input CNN-BiLSTM convolutional neural network takes relative position deviation data and multi-dimensional working condition parameters as input, and regresses and outputs PID control parameters in real time to realize fuzzy adaptive PID control to adapt to the dynamic changes in welding conditions.

[0019] The beneficial effects of this invention are: (1) This invention adopts a two-level enhancement architecture with the MSRCR enhancement submodule and the lightweight DeblurGAN-v2 correction submodule cascaded. The illumination enhancement is completed by the multi-scale Retinex algorithm with color restoration to eliminate the image degradation caused by smoke and splashes. The lightweight DeblurGAN-v2 generative adversarial network is used to repair the halo and background noise at the edge of the spot and preserve the gradient features of the weld edge. The lightweight DeblurGAN-v2 generative adversarial network adopts the WGAN-GP loss function with gradient penalty, with MobileNet-DSC as the backbone network and after model pruning and output quantization, so as to control the computation delay while ensuring the enhancement effect and adapt to the real-time processing requirements of laser welding. (2) The present invention adopts an improved lightweight CNN recognition module. By optimizing the network structure through hierarchical differentiated lightweight configuration, it can achieve pixel-level recognition and positioning of the laser spot center, spot area and weld edge, and stably output the relative position deviation data of the laser spot center relative to the weld. The network architecture can ensure the positioning accuracy of small gap welds while compressing the computation volume and improving the inference efficiency. It can simultaneously meet the real-time requirements of high-speed welding scenarios and the accuracy requirements of precision welding scenarios. After optimization, the single-frame inference delay is ≤10ms. (3) The intelligent closed-loop control unit of the present invention has a built-in Bayesian optimized multi-input CNN-BiLSTM convolutional neural network, which can integrate the relative position deviation data of visual feedback with welding material parameters, welding speed parameters, laser power parameters and other multi-dimensional real-time welding condition information to complete fuzzy closed-loop control calculation and dynamically output PID control parameters, realize adaptive adjustment of control parameters under different welding conditions, and effectively improve the response accuracy and operation stability of closed-loop control under dynamic fluctuations of working conditions. (4) The units of this invention communicate sequentially to form a complete intelligent closed-loop control link, realizing the information connection and collaborative operation of optical sensing, image enhancement, target recognition, intelligent control and execution feedback, forming a continuous dynamic closed-loop control; the system is equipped with a status monitoring and fault response mechanism, which can trigger an alarm and suspend welding operation when the relative position deviation data exceeds the set threshold or the execution unit fails, thereby improving the reliability of system operation and ultimately achieving a comprehensive improvement in welding quality consistency and production efficiency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall mechanical structure of the laser welding target recognition and intelligent closed-loop control system of the present invention; Figure 2 This is a flowchart illustrating the overall closed-loop workflow of the laser welding target recognition and intelligent closed-loop control method of the present invention. Figure 3 This is a flowchart illustrating the detailed execution process of the image intelligent enhancement preprocessing stage in this invention. Figure 4 This is a schematic diagram of the model architecture of the Bayesian optimized multi-input CNN-BiLSTM convolutional neural network used in this invention; Figure 5 This is a schematic diagram showing the effect comparison of the step-by-step image enhancement processing of the present invention; Explanation of reference numerals in the attached figures: 1. Specific wavelength oblique illumination source; 2. Coaxial laser welding vision perception unit; 3. Execution unit; 4. Welding laser head; 5. X-axis motion assembly; 6. Y-axis motion assembly; 7. Servo motor; 8. Lead screw slide. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] In the description of the invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention.

[0023] Example 1 like Figure 1As shown, the present invention provides a laser welding target recognition and intelligent closed-loop control system, including a specific wavelength oblique illumination source 1, a coaxial laser welding visual perception unit 2, an intelligent image enhancement and CNN target recognition unit, an intelligent closed-loop control unit, and an execution unit 3; each unit is sequentially connected in communication to form a complete intelligent closed-loop control link; The specific wavelength oblique illumination source 1 is made of a single wavelength LED that is different from the wavelength of the welding laser. It is obliquely arranged around the welding laser head 4, with the light emission direction facing the weld area of ​​the workpiece. It is used to provide auxiliary illumination to the weld area and provide basic lighting conditions for visual image acquisition. The coaxial laser welding vision sensing unit 2 is coaxially mounted on the reserved camera interface of the welding laser head 4, including an industrial lens, a narrowband filter, and a vision sensor arranged sequentially. The light transmission wavelength of the narrowband filter matches the emission wavelength of the specific wavelength oblique illumination source 1, which is used to filter out the strong arc light and plasma interference generated by the welding laser and improve the signal-to-noise ratio of the acquired image. At the same time, it can block the welding laser from directly shining on the vision sensor, forming optical protection for the sensing element. The vision sensor is used to acquire the laser spot and weld target image during the laser welding process and transmit the target image to the intelligent image enhancement and CNN target recognition unit. The intelligent image enhancement and CNN target recognition unit is used to perform intelligent enhancement preprocessing and pixel-level target recognition and localization on the acquired images in sequence, and output the relative position deviation data between the laser spot and the weld seam; the intelligent image enhancement and CNN target recognition unit includes a laser welding image intelligent enhancement module and an improved lightweight CNN recognition module connected in sequence. The intelligent enhancement module for laser welding images is used to perform dehazing, noise reduction, and feature enhancement processing on the input image to highlight the laser spot and weld edge features. The intelligent enhancement module for laser welding images includes an MSRCR enhancement submodule and a lightweight DeblurGAN-v2 correction submodule connected in sequence. The MSRCR enhancement submodule uses a multi-scale Retinex algorithm with color restoration to perform illumination enhancement processing on the input image and eliminate image degradation caused by smoke and spatter during the welding process. Furthermore, considering the characteristics of laser welding scenarios, such as localized high brightness, large background contrast, and interference from smoke and spatter, the multi-scale Gaussian standard deviation and gain coefficient of the multi-scale Retinex algorithm with color restoration are optimized in a targeted manner. The difference between the standard deviations of different scales is set to 3 to 4 times, and a lower gain coefficient is used to improve the brightness of the background area while avoiding excessive brightness in the laser spot area and its surroundings. The lightweight DeblurGAN-v2 correction submodule is used to eliminate image noise introduced after enhancement by the MSRCR algorithm, as well as edge halos in the strong light areas around the laser spot, and output the final enhanced image; The lightweight DeblurGAN-v2 correction submodule uses the WGAN-GP loss function with gradient penalty in its generative adversarial network, and its backbone network is replaced with MobileNet-DSC. The model is pruned and the output is quantized in advance to reduce the amount of computation and shorten the processing time, so as to meet the real-time requirements of the laser welding process. The lightweight DeblurGAN-v2 correction submodule uses laser welding images with halo and noise after being processed by the MSRCR algorithm as training samples to specifically repair the halo and background noise defects at the edge of the spot introduced by the MSRCR algorithm enhancement processing, while preserving the gradient features of the weld edge. The improved lightweight CNN recognition module is used to perform pixel-level recognition and localization of weld edges and laser spots in enhanced images. Its recognition targets include the center of the laser spot, the spot area, and the weld edge, and it outputs the relative positional deviation data of the laser spot center relative to the weld. The improved lightweight CNN recognition module is based on YOLOv5 and adopts a lightweight configuration: its backbone network is replaced with MobileNetV2, and it is equipped with quantization processing to compress the model size and computational load, thereby improving inference speed while ensuring recognition and localization accuracy, and adapting to the real-time requirements of welding scenarios.

[0024] The processing effects of the above two-level image enhancement can be found in [reference needed]. Figure 5 From left to right, the images are: the original welding image acquired by the coaxial laser welding visual perception unit 2; the preliminary enhanced image obtained by the MSRCR enhancement submodule; the final image after noise reduction and halo removal by the lightweight DeblurGAN-v2 correction submodule; and the target recognition and annotation completed by the improved lightweight CNN recognition module. The images can intuitively demonstrate the highlighting effect of weld edge features and the suppression effect of halo around the spot.

[0025] Furthermore, to address the need for precise position identification when weld gaps are less than 0.1mm, standard convolutions in the connection layer and detection head are retained to maintain detection accuracy; the model is quantized using INT8 to further improve detection inference speed, ultimately outputting pixel-level relative position deviation data; the improved lightweight CNN recognition module is trained using a multi-condition laser welding dataset of over 100,000 frames, and after model pruning and INT8 quantization lightweighting, the single-frame inference latency is ≤10ms.

[0026] The intelligent closed-loop control unit incorporates a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network. This network combines visual feedback relative position deviation data with real-time welding condition information to perform fuzzy closed-loop control calculations, outputting control signals to execution unit 3 and iteratively optimizing control parameters based on real-time welding data. The intelligent closed-loop control unit includes an adaptive PID controller. The inputs to the Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network include visual feedback relative position deviation data, welding material parameters, welding speed parameters, and laser power parameters. The output is the PID control parameters, used to achieve adaptive adjustment of control parameters under different welding conditions. The overall architecture of the aforementioned convolutional neural network can be found in [reference needed]. Figure 4 The main body of the network consists of a CNN feature extraction layer, a BiLSTM temporal modeling layer, and a connected output layer, which perform spatiotemporal feature fusion on the input multi-source data and output control parameters. The supporting Bayesian optimization module iteratively optimizes the network hyperparameters through a Gaussian process surrogate model and acquisition function until the prediction accuracy meets the set threshold, ensuring control accuracy and stability under all working conditions.

[0027] The execution unit 3 includes an X-axis motion component 5 and a Y-axis motion component 6 that are perpendicular to each other. Both the X-axis motion component 5 and the Y-axis motion component 6 include a servo motor 7 and a lead screw slide 8. The execution unit 3 is used to adjust the spatial position of the welding laser head 4 in real time according to the control signal so that the laser beam is always aligned with the center of the weld and realizes automatic tracking control of the weld.

[0028] Example 2 like Figure 2 As shown, the present invention also discloses a laser welding target recognition and intelligent closed-loop control method, which adopts the laser welding target recognition and intelligent closed-loop control system of Embodiment 1 above, including the following steps executed cyclically: S1. A specific wavelength oblique illumination source 1 is used to illuminate the weld area of ​​the workpiece at an oblique angle. The image is acquired by a coaxially mounted laser welding vision perception unit and narrowband optical filtering is performed to obtain a laser spot and weld target image that suppresses the strong arc light and plasma interference of the welding laser. S2. Based on the laser spot and weld target image obtained in step S1, the laser spot and weld target image are sequentially processed by intelligent enhancement preprocessing and improved lightweight convolutional neural network recognition and localization processing to obtain pixel-level relative position deviation data between the laser spot and the weld. like Figure 3 As shown, the intelligent enhancement preprocessing in step S2 specifically includes the following sub-steps: S21. Based on the input target image, the illumination enhancement process is performed using the multi-scale Retinex algorithm with color restoration to obtain a preliminary enhanced image that eliminates the degrading effects of smoke and splashes. S22. Based on the preliminary enhanced image obtained in step S21, noise and halo removal processing is performed using a lightweight DeblurGAN-v2 generative adversarial network to obtain the final enhanced image; The lightweight DeblurGAN-v2 generative adversarial network adopts the WGAN-GP loss function with gradient penalty, and its backbone network is replaced with MobileNet-DSC. It has also undergone model pruning and output quantization in advance to reduce computational latency and meet the real-time requirements of the laser welding process. S3. Based on the pixel-level relative position deviation data obtained in step S2, and combined with the real-time welding condition information of welding material, welding speed, and laser power, an adaptive fuzzy closed-loop control operation is performed through a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network to generate corresponding control signals and output them to the execution unit 3, driving the execution unit 3 to adjust the spatial position of the welding laser head 4. The Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network takes the relative position deviation data and multi-dimensional condition parameters as inputs, and regresses the output PID control parameters in real time to realize fuzzy adaptive PID control to adapt to the dynamic changes in welding conditions. S4. Real-time acquisition of relative position deviation data and action signals of execution unit 3 during the welding process and status monitoring and judgment; when the relative position deviation data exceeds the set threshold or execution unit 3 malfunctions, an alarm is triggered and the welding operation is suspended; under normal working conditions, the real-time operating data is fed back to the acquisition and control link, and the execution step S1 is returned to form a continuous dynamic closed-loop control.

[0029] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A laser welding target recognition and intelligent closed-loop control system, characterized in that, It includes a specific wavelength oblique illumination source, a coaxial laser welding visual perception unit, an intelligent image enhancement and CNN target recognition unit, an intelligent closed-loop control unit, and an execution unit; The specific wavelength oblique illumination source is made of a single-wavelength LED that is different from the wavelength of the welding laser, and is used to provide auxiliary illumination to the weld area; The coaxial laser welding vision perception unit is coaxially mounted on the reserved camera interface of the welding laser head, and includes an industrial lens, a narrowband filter, and a vision sensor arranged in sequence. The light transmission wavelength of the narrowband filter matches the emission wavelength of the specific wavelength oblique illumination source, and is used to filter out the strong arc light and plasma interference generated by the welding laser. The vision sensor is used to acquire the laser spot and weld target image during the laser welding process, and transmit the target image to the intelligent image enhancement and CNN target recognition unit. The intelligent image enhancement and CNN target recognition unit is used to perform intelligent enhancement preprocessing and pixel-level target recognition and positioning on the acquired image in sequence, and output the relative position deviation data between the laser spot and the weld seam. The intelligent closed-loop control unit incorporates a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network, which combines the relative position deviation data from visual feedback with real-time welding condition information to perform fuzzy closed-loop control calculations, outputs control signals to the execution unit, and iteratively optimizes control parameters based on real-time welding data. The execution unit includes a perpendicular X-axis motion component and a Y-axis motion component. Both the X-axis motion component and the Y-axis motion component include a servo motor and a lead screw slide. The execution unit is used to adjust the spatial position of the welding laser head in real time according to the control signal so that the laser beam is always aligned with the center of the weld.

2. The laser welding target recognition and intelligent closed-loop control system according to claim 1, characterized in that, The intelligent image enhancement and CNN target recognition unit includes a laser welding image intelligent enhancement module and an improved lightweight CNN recognition module connected in sequence. The laser welding image intelligent enhancement module is used to perform defogging, noise reduction and feature enhancement processing on the input image to highlight the laser spot and weld edge features; The improved lightweight CNN recognition module is used to perform pixel-level recognition and positioning of weld edges and laser spots in the enhanced image. Its recognition targets include the center of the laser spot, the spot area and the edge of the weld, and outputs the relative position deviation data of the center of the laser spot relative to the weld.

3. The laser welding target recognition and intelligent closed-loop control system according to claim 2, characterized in that, The laser welding image intelligent enhancement module includes an MSRCR enhancement submodule and a lightweight DeblurGAN-v2 correction submodule connected in sequence. The MSRCR enhancement submodule employs a multi-scale Retinex algorithm with color restoration to enhance the illumination of the input image and eliminate image degradation caused by smoke and spatter during welding. The lightweight DeblurGAN-v2 correction submodule is used to eliminate image noise introduced after enhancement processing by the MSRCR algorithm, as well as edge halos in the strong light area around the laser spot, and output the final enhanced image.

4. The laser welding target recognition and intelligent closed-loop control system according to claim 3, characterized in that, The lightweight DeblurGAN-v2 correction submodule's generative adversarial network adopts the WGAN-GP loss function with gradient penalty, and its backbone network is replaced with MobileNet-DSC. It has also undergone model pruning and output quantization in advance to reduce computation and shorten processing latency, thus meeting the real-time requirements of the laser welding process.

5. The laser welding target recognition and intelligent closed-loop control system according to claim 4, characterized in that, The lightweight DeblurGAN-v2 correction submodule uses laser welding images containing halos and noise, processed by the MSRCR algorithm, as training samples to specifically repair the halos and background noise defects at the edge of the spot introduced by the MSRCR algorithm enhancement processing, while preserving the gradient features of the weld edge.

6. The laser welding target recognition and intelligent closed-loop control system according to claim 1, characterized in that, The intelligent closed-loop control unit includes an adaptive PID controller; the inputs of the Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network include relative position deviation data from visual feedback, welding material parameters, welding speed parameters, and laser power parameters, and the output is PID control parameters, which are used to achieve adaptive adjustment of control parameters under different welding conditions.

7. A laser welding target recognition and intelligent closed-loop control method, characterized in that, The laser welding target recognition and intelligent closed-loop control system according to claim 1 includes the following steps executed cyclically: S1. A specific wavelength oblique illumination source is used to illuminate the weld area of ​​the workpiece at an oblique angle. The image is acquired by a coaxially mounted laser welding vision perception unit and narrowband optical filtering is performed to obtain a laser spot and weld target image that suppresses the strong arc light and plasma interference of the welding laser. S2. Based on the laser spot and weld target image obtained in step S1, the laser spot and weld target image are sequentially processed by intelligent enhancement preprocessing and improved lightweight convolutional neural network recognition and localization processing to obtain pixel-level relative position deviation data between the laser spot and the weld. S3. Based on the pixel-level relative position deviation data obtained in step S2, combined with the real-time welding condition information of welding material, welding speed, and laser power, adaptive fuzzy closed-loop control calculation is performed through a Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network to generate corresponding control signals and output them to the execution unit to drive the execution unit to adjust the spatial position of the welding laser head. S4. Real-time acquisition of relative position deviation data and execution unit action signals during the welding process and status monitoring and judgment; when the relative position deviation data exceeds the set threshold or the execution unit malfunctions, an alarm is triggered and the welding operation is suspended; under normal working conditions, the real-time operating data is fed back to the acquisition and control link, and the execution step S1 is returned to form a continuous dynamic closed-loop control.

8. The laser welding target recognition and intelligent closed-loop control method according to claim 7, characterized in that, The intelligent enhancement preprocessing described in step S2 specifically includes the following sub-steps: S21. Based on the input target image, the illumination enhancement process is performed using the multi-scale Retinex algorithm with color restoration to obtain a preliminary enhanced image that eliminates the degrading effects of smoke and splashes. S22. Based on the preliminary enhanced image obtained in step S21, noise and halo removal processing is performed using a lightweight DeblurGAN-v2 generative adversarial network to obtain the final enhanced image.

9. The laser welding target recognition and intelligent closed-loop control method according to claim 8, characterized in that, The lightweight DeblurGAN-v2 generative adversarial network mentioned in step S22 adopts the WGAN-GP loss function with gradient penalty, its backbone network is replaced with MobileNet-DSC, and it has been pre-processed with model pruning and output quantization to reduce computational latency and meet the real-time requirements of the laser welding process.

10. The laser welding target recognition and intelligent closed-loop control method according to claim 7, characterized in that, In step S3, the Bayesian-optimized multi-input CNN-BiLSTM convolutional neural network takes relative position deviation data and multi-dimensional working condition parameters as inputs, and regresses the output PID control parameters in real time to realize fuzzy adaptive PID control, so as to adapt to the dynamic changes of welding conditions.

Images