An automated welding method for steel cage welding robots

Abstract

This invention discloses an automatic welding method for a rebar cage welding robot, relating to the field of welding robot technology. Based on an improved ResNet-50 multimodal vision system, the welding point recognition error is reduced to the 0.1mm level. Combined with the SE(3) spatial coordinate transformation algorithm, the positioning accuracy reaches ±0.03mm (3σ). The vision algorithm, which integrates deformable convolution and coordinate attention mechanisms, still maintains a 23.6% improvement in recognition rate under illumination fluctuations of 200-1500 lux. The improved RRT algorithm, combined with adaptive B-spline interpolation, improves the motion smoothness of the six-degree-of-freedom robotic arm by 400%, effectively avoiding welding defects caused by abrupt path changes. The motion control of the robotic arm is optimized through a friction compensation model, significantly improving welding stability.

Description

Technical Field

[0001] This invention relates to the field of welding robot technology, and in particular to an automatic welding method for a rebar cage welding robot. Background Technology

[0002] With the acceleration of industrialization in construction, the quality of steel reinforcement cages, as the key skeleton of concrete structures, directly affects the safety of engineering projects.

[0003] Current steel cage manufacturing primarily utilizes semi-automatic roll welding equipment. While this equipment can automatically transport main reinforcement bars and stirrups, significant technical bottlenecks remain in the core welding process: First, welding point positioning relies on manual visual inspection, with an error rate as high as 15% under complex lighting conditions, leading to frequent instances of incomplete or missed welds. Second, traditional path planning algorithms cannot adapt to the three-dimensional spatial structure of the steel cage, resulting in abrupt changes in the robotic arm's trajectory, causing welding spatter and deformation. These problems not only cause material waste rates of 3%-5% but also result in labor costs exceeding 40% of the total cost, severely hindering the construction progress of large-scale infrastructure projects.

[0004] In response to the above phenomena, the development of an efficient and precise automated welding technology is particularly urgent. Therefore, an automated welding method for steel cage welding robots is proposed to meet the needs of practical applications. Summary of the Invention

[0005] This invention provides an automatic welding method for steel cage welding robots, which solves the technical problems of low automation and poor adaptability of traditional steel cage welding equipment.

[0006] To address the related technical problems, this invention provides an automated welding method for steel cage welding robots, comprising the following steps:

[0007] Step S1: The main reinforcing bars are automatically fed by the storage mechanism, and the steel bars are cut to length and threaded through the roller conveyor system. The main reinforcing bars are positioned, aligned and clamped by the transmission device.

[0008] Step S2: After each main rib is delivered, the automatic rotating mechanism of the skeleton drives the servo motor to rotate the internal long shaft by a preset angle, so that the stirrup support device reaches the next material placement station.

[0009] Step S3: After the main reinforcement and stirrups are laid out, the vision system captures the image of the welding area through imaging technology, generates a probability heat map of the welding point, and obtains the three-dimensional coordinates of the welding point through SE(3) spatial coordinate transformation.

[0010] In step S4, the six-axis robot, which is inverted and mounted on the cantilever structure, plans a collision-free path in the three-dimensional work space according to the spatial distribution of the welding task, and generates a smooth welding trajectory through adaptive B-spline interpolation. During the trajectory execution, a friction compensation model is introduced to suppress nonlinear friction interference, ensuring accurate spatial coverage of the welding points and process stability, and finally completing the skeleton welding.

[0011] In some embodiments, step S3, generating a probability heatmap of weld points, includes the following sub-steps:

[0012] Step S301: Preprocess the original image of the welding area to obtain a 256-level edge intensity map and semantic features extracted by the improved ResNet-50.

[0013] Step S302: The preprocessed 256-level edge intensity map and the semantic features extracted by the improved ResNet-50 are input into the feature pyramid network for channel concatenation.

[0014] Step S303: Cross-modal feature interaction is achieved by using a 3×3 convolution kernel on the concatenated features to generate a fused feature map of spatial and semantic domains.

[0015] Step S304: The fused feature map is upsampled to the original image size using a bilinear interpolation algorithm to obtain a welding point probability heatmap with sub-pixel accuracy.

[0016] In some embodiments, the preprocessing process in step S301 includes:

[0017] Step S3011: An edge feature map is generated using the Canny edge detection algorithm with adaptive gamma correction, wherein the function for adaptive gamma correction is:

[0018] γ(x,y)=γ0+k·I(x,y),

[0019] Where: γ0=0.4, j=0.1, I(x,y) is the mean value of all pixels within a 5×5 window centered at (x,y);

[0020] Step S3012: The edge feature map is processed based on the CLAHE algorithm. An adaptive local histogram equalization method with a fixed block size of 32×32 is adopted, and a brightness-sensitive limiting mechanism is introduced to generate a 256-level edge intensity map. The limiting threshold is:

[0021]

[0022] Where: μ block μ is the mean of the local block centered at pixel (x,y). imageT is the mean of the entire image. max The maximum amplitude limit threshold;

[0023] Step S3013: Extract semantic features of the original image using the improved ResNet-50.

[0024] In some embodiments, the semantic feature extraction process of the improved ResNet-50 in step S3013 is as follows:

[0025] In the third and fourth residual blocks of ResNet-50, deformable convolutional layers are embedded after the 3×3 convolutional layers in each residual block to adapt to the deformation characteristics of the weld points.

[0026] After the residual block output, a coordinate attention module is cascaded to perform global average pooling on the width and height directions of the original image to enhance the geometric features of the welding points.

[0027] In some embodiments, the global average pooling step specifically includes:

[0028] Step S30131, Multi-dimensional spatial pooling:

[0029] For the feature map of each channel, the average values ​​are taken in the width and height directions respectively to obtain tensors with shapes (C,1,W) and (C,H,1) respectively;

[0030] Step S30132, cross-channel weight generation;

[0031] Generate response weights for specific spatial dimensions respectively:

[0032] W x =σ(CN(AvgPool(F;x))),

[0033] W y =σ(CN(AvgPool(F;y))),

[0034] Where CN is a 1×1 convolution used to integrate information across channels and generate weights, and σ is a ReLU activation function that introduces nonlinearity.

[0035] Step S30133, Fusion Modulation:

[0036] First, W x and W y Broadcast to the same spatial dimension ((C,H,W)) as the input feature F to obtain the comprehensive weight map W. x +W y ;

[0037] Then input the feature F and multiply it element-wise with the weight map:

[0038] F′=F×(Wx +W y ),

[0039] The edges of the weld points are enhanced, while the background area is suppressed;

[0040] Step S30134, residual connection output:

[0041] The residual path of the original input F is retained, and the final output is:

[0042] Output = F + F' = F + F × (W) x +W y ).

[0043] In some embodiments, step S3, the SE(3) spatial coordinate transformation specifically includes the following steps:

[0044] Step S305, establish the closed-loop transformation matrix from the image coordinate system to the tool coordinate system based on Eye-in-Hand calibration, specifically including the following sub-steps:

[0045] Step S3051: At least 6 different viewpoints of images and pose data are collected by using a high-contrast checkerboard calibration board in conjunction with multi-pose sampling of a six-axis robot, and the homography matrix is ​​solved by the DLT algorithm to construct a mapping model.

[0046] Step S3052: The initial values ​​of the camera extrinsic parameters are decomposed using the Tsai-Lenz radial constraint method, and the rotation matrix R and translation vector t are iteratively optimized using the Levenberg-Marquardt algorithm to minimize the projection error function.

[0047]

[0048] Among them, P i To calibrate the world coordinates of the corner points of the plate, p i Here, Proj represents the image coordinates, and Proj is the camera intrinsic projection function, until the error change is less than 1×10. -6 ;

[0049] Step S3053: Combining the pseudo-inverse solution strategy, the optimized camera pose is... With the robot's end-effector pose Substitute into the closed-loop transformation matrix:

[0050]

[0051] Expanding into a system of linear equations Ax = b, the pseudoinverse A is obtained by Tikhonov regularization of the coefficient matrix A and singular value decomposition. + The least squares solution is obtained:

[0052] x = A+ ·b;

[0053] Step S3054 verifies consistency through repeated random posture tests to ensure that the three-dimensional coordinate transformation accuracy of the welding point meets the sub-millimeter positioning requirements.

[0054] In some embodiments, the improved RRT algorithm in step S4 includes:

[0055] Step S401: Define a three-dimensional search domain containing dynamic obstacles in the robot's workspace;

[0056] Step S402: Node expansion is performed using an adaptive step size factor α, with an initial value of 1% to 3% of the maximum path length; and α is dynamically adjusted based on the Euclidean distance d, local space complexity ρ, and adjustment factor k.

[0057]

[0058] Step S403 extends the method of using bounding box-triangular mesh cross detection to determine the interference between the path segment and the steel cage structure, and introduces a dynamic obstacle time series prediction model to evaluate the overlapping area of ​​the motion trajectory, thus ensuring path safety.

[0059] In some embodiments, in step S4, the adaptive B-spline interpolation process dynamically sets the control point density and node vector spacing based on the spatial angle and curvature between welding points, generating values ​​that satisfy velocity ≤ 0.5 m / s and acceleration ≤ 3 m / s². 2 Constrained continuous trajectory.

[0060] In some embodiments, the friction compensation model in step S4 is:

[0061]

[0062] Where, τ cmd This is the theoretical torque command; Joint angular velocity; For symbolic functions, and These are the viscous friction coefficient and the Coulomb friction coefficient identified offline, respectively.

[0063] In some embodiments, in step S2, the major axis rotates by a preset angle:

[0064] θ = 360° / N

[0065] Where N represents the total number of circumferential main reinforcement bars, compared with related technologies, the automatic welding method for steel cage welding robots provided by this invention has the following beneficial effects:

[0066] This invention provides an automated welding method for rebar cage welding robots. Based on an improved ResNet-50 multimodal vision system, the welding point recognition error is reduced to the 0.1mm level. Combined with the SE(3) spatial coordinate transformation algorithm, the positioning accuracy reaches ±0.03mm (3σ). The vision algorithm, which integrates deformable convolution and coordinate attention mechanisms, maintains a 23.6% improvement in recognition rate even under illumination fluctuations of 200-1500 lux. The improved RRT algorithm, combined with adaptive B-spline interpolation, improves the motion smoothness of the six-degree-of-freedom robotic arm by 400%, effectively avoiding welding defects caused by abrupt path changes. The motion control of the robotic arm is optimized through a friction compensation model, significantly improving welding stability. Practical applications show that this method saves 60% of labor costs, reduces material loss by 3%, and increases single-shift production capacity by 4 times compared to traditional processes, providing a reliable automated manufacturing solution for rebar cages in major projects such as bridges and tunnels. Attached Figure Description

[0067] Figure 1 This is a flowchart of the automatic welding process of the present invention;

[0068] Figure 2 This is a schematic diagram of the overall device structure of the present invention;

[0069] Figure 3 This is a schematic diagram of the material storage mechanism of the present invention;

[0070] Figure 4 This is a schematic diagram of the automatic frame rotation mechanism of the present invention;

[0071] Figure 5 This is a schematic diagram of the six-axis robot structure of the present invention.

[0072] The following are the labeling elements in the diagram: 100, material storage mechanism; 101, material storage rack; 102, connecting angle steel; 200, roller conveyor system; 201, sawing and threading device; 202, transmission device; 300, automatic frame rotation mechanism; 301, long shaft; 302, stirrup support device; 303, bottom beam support mechanism; 400, six-axis robot; 500, cage traveling trolley. Detailed Implementation

[0073] Example 1

[0074] This embodiment provides an automated welding method for steel cage welding robots, such as... Figure 1 As shown, the present invention includes the following steps:

[0075] S1, the main rib is automatically fed by the storage mechanism 100, and the fixed-length cutting and thread processing are completed by the roller system 200 with double-end sawing and threading device 201. The positioning, alignment and clamping of the main rib are realized by the transmission device 202.

[0076] S2, after each main rib is delivered, the skeleton automatic rotation mechanism 300 drives the servo motor to rotate the internal long shaft 301 by a preset angle θ = 360° / N, where N is the total number of circumferential main ribs, so that the stirrup support device 302 reaches the next material placement station.

[0077] S3, after the main reinforcement and stirrups are laid out, the vision system captures the image of the welding area through imaging technology, generates a probability heat map of the welding point, and obtains the three-dimensional coordinates of the welding point through SE(3) spatial coordinate transformation;

[0078] S4, the six-axis robot 400, which is inverted and mounted on the cantilever structure, plans a collision-free path in the three-dimensional work space according to the spatial distribution of the welding task, and generates a smooth welding trajectory through adaptive B-spline interpolation. During the trajectory execution, a friction compensation model is introduced to suppress nonlinear friction interference, ensuring accurate spatial coverage of the welding points and process stability, and finally completing the skeleton welding.

[0079] The following is a detailed explanation of the overall process of the automatic welding method for rebar cage welding robots:

[0080] In step S1, the main rib is first placed on the storage mechanism 100, and then attached... Figure 2 The system consists of four sets of single-piece storage racks 101, with the bottom bolted to the entire base beam. Two sets of symmetrical connecting angle steels 102 are then bolted together to secure the storage racks 101 into a single unit, supporting the raw material, reinforcing bars. After the reinforcing bars are fed into the roller conveyor system 200 of the production line, they undergo automatic sawing and threading of the main reinforcing bars. Each end of the roller conveyor is equipped with a sawing and threading device 201 to enable rapid unloading of the reinforcing bars. This drives a chain drive to transport the reinforcing bars from the bottom to the top. Through a transmission device 202, the unloaded main reinforcing bars are automatically conveyed. A main reinforcing bar receiving device is also included to receive, position, align, and tighten the main reinforcing bars, facilitating integration with subsequent processes.

[0081] In step S2, the reinforcing bars are according to the attached... Figure 2 The structure operates via a bottom beam support mechanism 303, which includes a powered, rotatable long shaft 301 (driven by a servo motor). Six sets of stirrup support devices 302 are mounted on the long shaft 301 (this mechanism is pneumatically controlled and can extend or retract as needed to allow the equipment to be fitted or separated from the rebar cage). During production, the long shaft 301 and stirrup support devices 302 are located inside the rebar cage, providing support from the inside out. The rotation of the long shaft 301, in conjunction with the main reinforcement receiving device, enables the individual placement and welding of the main reinforcement bars and stirrups.

[0082] After the steel cage is welded, the bottom beam support mechanism 303 on both sides of the base frame is lifted as a whole, and then the stirrup support device 302 retracts to separate it from the steel cage. Then, the steel cage is sent to the next station by the alternating operation of the cage-carrying trolley 500, so that the steel cage is bent and separated from the internal support long axis 301.

[0083] In step S3, a six-axis robot 400 integrated with a vision system is used to automatically locate and weld the main reinforcement bars and stirrups of the steel cage. After the main reinforcement bars are arranged, the vision system captures images of the welding area using imaging technology and identifies the positions of the welding points using image recognition algorithms. Subsequently, the six-axis robot 400 navigates based on this image data, automatically moves to each welding point, and performs the welding operation. This step significantly improves the accuracy and efficiency of welding while reducing human intervention and ensuring consistent welding quality.

[0084] In step S4, according to the appendix Figure 5 The structure features a six-axis robot 400 suspended upside down on a cantilever, equipped with a vision system to achieve comprehensive coverage of all welding points in the rebar cage. This layout allows the six-axis robot 400 to flexibly reach every corner of the rebar cage and perform precise welding operations at each location.

[0085] This method allows the entire framework of the rebar cage to be welded completely and efficiently, ensuring the structural robustness and high-quality completion of the welding work. It overcomes the shortcomings of traditional manual methods, which struggle to fundamentally improve quality weaknesses. It saves 60% on labor costs, approximately 3% on material costs, increases efficiency by 400% compared to manual welding, improves the recognition rate by 23.6%, effectively supports precision welding operations down to 0.1mm, greatly improves the quality and efficiency of rebar cage welding, shortens processing time, and guarantees product quality.

[0086] Example 2

[0087] Based on Embodiment 1, this embodiment provides a production line using an automatic welding method for steel cage welding robots, including a material storage mechanism 100, a roller conveyor system 200, an automatic frame rotation mechanism 300, a six-axis robot 400, and a cage-carrying trolley 500.

[0088] The storage mechanism 100 consists of four sets of single-piece storage racks 101 and two sets of symmetrical connecting angle steels 102. The bottom of the four sets of single-piece storage racks 101 is fixed to the entire bottom beam with bolts. The two sets of connecting angle steels 102 are fixed to the storage racks 101 as a whole by bolts to support the raw material steel bars.

[0089] The roller conveyor system 200 includes two sets of sawing and wire-threading devices 201 and a conveying device 202. The two sets of sawing and wire-threading devices 201 are respectively set at both ends of the conveying device 202. After the two ends of the main rib fed into the conveying device 202 are sawed and wire-threaded by the two sets of sawing and wire-threading devices 201, they are sent into the automatic frame rotation mechanism 300 through the conveying device 202.

[0090] The automatic rotation mechanism 300 for the steel cage includes a long shaft 301, six sets of stirrup support devices 302, and a bottom beam support mechanism 303. The six sets of stirrup support devices 302 are arranged on the surface of the long shaft 301, which is driven by a servo motor through a reducer. The bottom beam support mechanism 303 is arranged below the long shaft 301 to support the steel cage. During production, the long shaft 301 and the stirrup support devices 302 are located inside the steel cage, providing support to the steel cage from the inside out. Through the rotation of the long shaft 301 and in coordination with the main reinforcement receiving device, the main reinforcement and stirrups are laid and welded one by one.

[0091] Two six-axis robots 400 are installed, located above both ends of the automatic frame rotation mechanism 300, working together. The front end of each six-axis robot 400 is equipped with a resistance welding device, which welds and fixes the main reinforcing bars and stirrups. A cage-carrying trolley 500 is positioned at the bottom of the automatic frame rotation mechanism 300. After welding, the stirrup support device 302 is released, separating the automatic frame rotation mechanism 300 from the reinforcing cage frame. The reinforcing cage frame then rests on the cage-carrying trolley 500. The welded reinforcing cage frame is moved by controlling the movement of the cage-carrying trolley 500.

[0092] Example 3

[0093] Building upon Example 1, this example utilizes a convolutional neural network (CNN), combined with edge detection algorithms (such as the Canny operator), and a feature extraction network based on an improved ResNet-50 framework to accurately locate welding points in images, thereby enhancing the ability to extract weld edge features. To address issues related to lighting variations and background interference, robustness is enhanced through multimodal feature fusion.

[0094] Specifically, the original image of the welding area is first preprocessed. The preprocessed 256-level edge intensity map and the semantic features extracted by the improved ResNet-50 are then input into a feature pyramid network for channel concatenation. Next, a 3×3 convolutional kernel is used for cross-modal interaction to generate a feature map containing both spatial and semantic information. A bilinear interpolation algorithm is then used to upsample the fused feature map, ensuring its output size matches the original input image, ultimately yielding a sub-pixel precision probability heatmap of the welding points.

[0095] Furthermore, an adaptive gamma-corrected Canny edge detection algorithm is used to generate edge feature maps, which are then processed using the CLAHE algorithm (contrast-limited adaptive histogram equalization). Semantic features of the original image are extracted using an improved ResNet-50. The preprocessed 256-level edge intensity map and the semantic features extracted by the improved ResNet-50 are then input into a feature pyramid network (FPN) for channel concatenation.

[0096] The adaptive gamma correction dynamically adjusts the enhancement curves for different regions based on local image brightness to improve the integrity and contrast of edge detection under uneven illumination conditions. The generated edge feature map provides stable geometric texture information. This is further processed using the CLAHE algorithm to enhance the image's brightness and structural features. A brightness-sensitive limiting mechanism is introduced to generate a 256-level edge intensity map, laying a high-quality input foundation for subsequent deep semantic feature extraction and multimodal fusion. The limiting coefficient is dynamically adjusted based on the ratio of the local region's average brightness to the overall image's average brightness, thereby improving the detail in dark areas and suppressing noise amplification in highlight areas.

[0097] Specifically, an adaptive local histogram equalization method with a fixed block size of 32×32 is adopted, and a brightness-sensitive limiting mechanism is introduced to generate a 256-level edge intensity map to prevent excessive magnification of details in low-light areas. The adaptive gamma correction function uses γ(x,y)=γ0+k·I(x,y) to perform brightness correction on the input image, where: γ0=0.4, γ0 is the base gamma value to ensure basic brightness adjustment capability in all areas; k=0.1, k is a sensitivity coefficient to control the intensity of the influence of local brightness on the gamma value; I(x,y) is the local brightness of pixel (x,y), which is obtained by calculating the mean value of all pixels within a 5×5 window centered at (x,y).

[0098] In this embodiment, the limiting threshold for the brightness-sensitive limiting mechanism introduced based on the CLAHE algorithm is:

[0099]

[0100] Where: μ block The mean value of a local block (32×32) centered at pixel (x,y) reflects local brightness; μ image T represents the average brightness of the entire image, reflecting the overall brightness. max =150, T max This is the maximum threshold value, used to control the upper limit of the dynamic threshold.

[0101] After introducing a brightness-sensitive limiting mechanism, the low-light region (μ block <μ image): Decreasing T(x,y) limits the contrast stretching range and avoids excessive noise amplification; Low-light region (μ block >μ image ): Increasing T(x,y) allows for stronger contrast stretching while preserving details in bright areas.

[0102] The image is divided into non-overlapping 32×32 blocks, and histogram equalization is performed independently on each block. The large block size covers a sufficient number of pixels, reducing the sparsity of local histograms (especially in low-texture areas). Furthermore, the fixed size avoids the complexity of adaptive block size and facilitates parallel computation.

[0103] In this embodiment, in the third and fourth residual blocks of ResNet-50, deformable convolutional layers are embedded after the 3×3 convolutional layers of each residual block to adapt to the deformation features of the welding points; and a coordinate attention module is cascaded after the output of the residual blocks to perform global average pooling on the width and height directions of the original image respectively, thereby enhancing the geometric features of the welding points, thus obtaining an improved ResNet-50.

[0104] In the third and fourth residual blocks of ResNet-50, deformable convolutional layers are embedded after the 3×3 convolutional layers in each residual block. Traditional convolutions have a fixed receptive field, while deformable convolutions dynamically adjust the sampling point positions to adapt to complex lighting conditions in industrial scenarios. The sampling points of each convolutional kernel dynamically shift according to the input features, thus capturing local features more flexibly. This improves adaptability to irregularly shaped welding points and alleviates feature misalignment problems caused by deformation.

[0105] In this embodiment, a coordinate attention mechanism is embedded after the 3×3 convolutional layer of each residual block, combined with the global average pooling channels in the width and height directions of the original image. Specifically:

[0106] For the feature map of each channel, the average values ​​are taken in the width and height directions respectively to obtain tensors with shapes (C,1,W) and (C,H,1) respectively;

[0107] Generate response weights W for a specific spatial dimension. x W y ,

[0108] Weight calculation: W x =σ(CN(AvgPool(F;x))), W y =σ(CB(AvgPool(F;y)))

[0109] Where CN is a 1×1 convolution used to integrate information across channels and generate weights, and σ is the ReLU activation function, which introduces nonlinearity.

[0110] W x and Wy Broadcast to the same spatial dimension ((C,H,W)) as the input feature F to obtain the comprehensive weight map W. x +W y .

[0111] Then perform feature modulation: multiply the input feature F element-wise with the weight map: F′=F×(W x +W y In this method, regions with high weights (such as the edges of weld points) are enhanced, while regions with low weights (such as the background) are suppressed.

[0112] Residual Connection Preservation: The residual paths of the original input F are preserved, and the final output is:

[0113] Output = F + F' = F + F × (W) x +W y )

[0114] Residual connections ensure that gradients flow directly back to the input, avoiding the vanishing gradient problem during deep network training.

[0115] All structural modifications retain the original residual connection paths, ensuring stable gradient propagation in deep networks and maintaining effective reuse between low-level details and high-level semantics, thereby improving the network's ability to perceive the structure and identify the location of complex lighting and deformed welds.

[0116] Specifically, the Layer 4 output of ResNet-50 (default 2048 channels, 7×7 spatial size) is adjusted from 2048 dimensions to 1024 dimensions through 1×1 convolution; then, bilinear interpolation is used to restore the 7×7 feature map to the same spatial size (H×W) as the Canny map.

[0117] Then, a 256-level edge intensity map (256 channels, H×W) and semantic features extracted by the improved ResNet-50 (1024 channels, H×W) are input, resulting in a 1280-channel (256+1024) cascaded feature map, where each spatial location contains information from two modalities. A 3×3 convolutional kernel is then input to mix the features from different modalities within a local spatial range, promoting the fusion of edge details and semantic information.

[0118] Among them, the channel cascading method can preserve the integrity of edge geometric features and deep semantic information to the greatest extent without destroying the original spatial layout, and build a cross-modal feature intersection interface in the convolutional neural network, providing a structural foundation for subsequent unified feature representation.

[0119] Deep fusion between geometric edge features and semantic context is achieved through local convolution operations. This process strengthens the structural correspondence between modalities based on shared convolution weights, improves the saliency response and spatial consistency of the weld target region in the joint feature map, and generates a fused feature map that contains spatial-semantic dual-domain representation capabilities.

[0120] The bilinear interpolation algorithm constructs a weighted average function among four neighboring points around the target pixel location to restore the continuous feature mapping of non-integer locations, preserving the spatial continuity and probability distribution smoothness of the heatmap response, and finally outputting a high-resolution probabilistic heatmap of the welding point.

[0121] In summary, the channel cascading and convolutional interaction mechanism based on low-level structural edge maps and high-level semantic features can organically combine local geometric details with overall target perception capabilities while maintaining the spatial consistency of the original image. This fusion strategy significantly improves the response intensity and location reliability of the weld area in the feature map, enhances the robustness of the model under complex lighting, weld point occlusion, deformation interference and other industrial field conditions, and has the advantages of clearer response and stronger background suppression in the generation of the final weld point probability heatmap, providing a high signal-to-noise ratio visual perception foundation for three-dimensional spatial coordinate transformation.

[0122] Example 4

[0123] Based on Examples 1 to 3, SE(3) represents a special Euclidean group in three-dimensional space, which includes rotation (SO(3)) and translation (R). 3 This combination of coordinates is used to describe rigid body motion (such as the transformation between the robot's end-effector coordinate system and the camera coordinate system). The goal is to construct an accurate closed-loop transformation matrix from the image coordinate system (Camera Frame) to the tool coordinate system (Tool Frame). Achieve sub-millimeter level positioning accuracy.

[0124] A closed-loop transformation matrix from the image coordinate system to the tool coordinate system was established through eye-in-hand calibration. A mapping model was constructed by using a calibration board in conjunction with multi-pose sampling of a six-axis robot. The Tsai-Lenz model was used for iterative optimization to minimize the projection error, and a pseudo-inverse solution strategy was combined to improve the solution stability under small sample conditions, with a calibration error ≤0.05mm.

[0125] The closed-loop transformation matrix is:

[0126]

[0127] Where W is the world coordinate system, which in this embodiment is the coordinate system where the calibration plate is located; The pose of the camera in the world coordinate system (solved using a calibration plate); The pose of the robot's end effector (obtained directly from the robot control system).

[0128] The specific steps are as follows:

[0129] (1) Calibration plate design and posture sampling: A high-contrast checkerboard calibration plate with known dimensions (e.g., 10×10cm) is used. The coordinates of the corner points / centers are pre-calibrated (under the world coordinate system W). A camera is mounted on the end effector of the six-axis robot. The tool coordinate system (e.g., welding torch tip) and the camera extrinsic parameters are fixed. The robot moves in at least 6 different postures, covering different perspectives of the calibration plate, and images and robot pose data (joint angles or end effector pose) are collected. ).

[0130] (2) Initial calibration and nonlinear optimization of the Tsai-Lenz model: The Harris corner detection algorithm was used for each image to extract the sub-pixel coordinates of the checkerboard corners. The homography matrix H of each image was solved using the DLT (Direct Linear Transformation) algorithm. Based on the Tsai-Lenz radial constraint method, the initial values ​​of the camera extrinsic parameters were obtained by decomposition. The initial value of R is extracted using the rotation part of H with an error ≤0.5°, and the initial value of t is extracted using the translation part with an error of 2mm.

[0131] Define the projection error function:

[0132]

[0133] Among them, P i To calibrate the world coordinates of the corner points of the plate, p i Here are the image coordinates, and Proj is the camera intrinsic projection function.

[0134] The Levenberg-Marquardt (LM) algorithm is used to iteratively optimize R and t until the error change is less than 1 × 10⁻⁶. -6 Or the number of iterations is ≥50.

[0135] (3) Pseudo-inverse solution: The optimized camera pose is obtained by solving the pseudo-inverse problem. and Substitute into the transformation matrix:

[0136]

[0137] Expanding this into a system of linear equations Ax = b, where x is the vector of transformation parameters to be determined.

[0138] The coefficient matrix A is subjected to Tikhonov regularization with a regularization parameter λ = 0.01 to suppress noise sensitivity.

[0139] Solving the pseudoinverse A using singular value decomposition +The least squares solution is obtained:

[0140] x = A + ·b

[0141] (4) Closed-loop verification and accuracy validation

[0142] The robot randomly selects three calibration poses and repeats the process five times, recording the consistency of the transformation between the tool coordinate system and the camera coordinate system.

[0143] Set the target coordinates of the welding point and align them using a visual positioning guide tool. Repeat the positioning process multiple times to ensure that the maximum deviation is ≤0.1mm.

[0144] The calibration error of ≤0.05mm is set based on the requirements of welding task for point spatial accuracy control. This value is half of the maximum allowable projection error required to achieve 0.1mm level welding point coverage accuracy, ensuring that the image spatial error will not be amplified into an unacceptable three-dimensional deviation after projection transformation.

[0145] Example 5

[0146] Based on embodiments one through four, the improved RRT algorithm in this embodiment includes:

[0147] Step S401: Define a three-dimensional search domain containing dynamic obstacles in the robot's workspace;

[0148] Step S402: Node expansion is performed using an adaptive step size factor α. Its initial value is set based on the minimum controllable motion resolution of the robot's end effector and the search space scale, typically ranging from 1% to 3% of the maximum path length. The adaptive adjustment rule dynamically corrects α based on the Euclidean distance d between the current expansion direction and the target point, and the local spatial complexity of the current node. Specifically, the following function is used:

[0149]

[0150] Where ρ represents obstacle density and k is an adjustment factor to ensure that the step size is automatically reduced in dense obstacle areas and automatically increased in open areas;

[0151] Step S403: Before each expansion of a new path segment, determine in real time whether there is geometric interference between the path segment and the steel cage structure and dynamic obstacles. Specifically, perform rapid collision detection using the bounding box-triangular mesh cross detection method. At the same time, introduce a time axis prediction model for dynamic obstacles, combine obstacle velocity vectors and motion trajectories to evaluate possible overlapping areas in the future, realize the feasibility determination of the dynamic path, and thus ensure the safety and accessibility of the final path during execution.

[0152] By using an adaptive step size factor, the convergence speed of path search can be significantly improved, and the problem of local oversampling can be effectively avoided, thereby improving path smoothness and planning efficiency.

[0153] Example 6

[0154] Based on Example 5, for the discrete welding path points generated by the improved RRT algorithm, the control point density and node vector spacing of the B-spline are dynamically set according to the spatial angle and curvature between the welding points to achieve fine-grained fitting of the local deformation of the path; in Cartesian space, points satisfying velocity ≤ 0.5 m / s and acceleration ≤ 3 m / s² are generated. 2 Constrained continuous trajectory.

[0155] The adaptive B-spline interpolation algorithm, based on the traditional B-spline path representation, combines the topological characteristics of the cylindrical skeleton of the reinforced cage structure and the distribution of continuous points with multiple curvatures. It introduces a trajectory local curvature feedback adjustment mechanism and a dynamic control point distribution strategy. Compared with ordinary B-spline interpolation methods, this adaptive strategy can spatially resolve and adjust according to the local structural complexity of the reinforced cage components, significantly improving the attitude stability and trajectory tracking accuracy of the robot's end-effector welding torch during continuous welding. This avoids problems such as uneven welds, incomplete welds, or over-welding caused by sudden path changes, ensuring consistent welding quality and dynamic stability of trajectory execution. Furthermore, a friction compensation model is used.

[0156]

[0157] Eliminate the nonlinear effects of robotic arm dynamics; where τ cmd The theoretical torque command output by the controller; Joint angular velocity; This is a sign function, and its value is +1 or -1 depending on the direction of joint movement. and These are the viscous friction coefficient and the Coulomb friction coefficient, obtained through offline identification, respectively. and The experiment involved applying a set of constant driving currents (or torques) of different magnitudes and directions to each joint of the robot and recording the steady-state angular velocity response curves of the joints under no-load conditions. Under constant temperature conditions, the velocity-torque data points during joint movement were fitted, establishing a linear fitting relationship with joint angular velocity as the horizontal axis and steady-state driving torque as the vertical axis. The slope of the fitted curve corresponds to the viscous friction coefficient. Intercept corresponding to Coulomb friction coefficient By performing regression analysis on the torque-velocity characteristic data using the least squares method, the parameters of the friction model for each joint were obtained, and finally, a nonlinear joint model suitable for friction compensation control was constructed.

[0158] Actual tests show that the system can achieve a welding point positioning accuracy of ±0.03mm (3σ) and a single point processing cycle of ≤85ms under conditions such as illuminance fluctuations (200-1500 lux) and splash interference. The recognition rate is 23.6% higher than that of traditional single algorithm solutions, effectively supporting precision welding operations at the 0.1mm level.

[0159] Through continuous iteration of the RRT algorithm, the six-axis robot 400 can plan a collision-free path in the complex steel cage welding environment, ensuring that the inverted six-axis robot 400 can effectively avoid obstacles and successfully complete the welding task of all welding points of the steel cage when moving flexibly in six degrees of freedom.

[0160] In practical applications, combined with other technologies of the vision-robot collaborative control system, such as precise welding point positioning, adaptive path planning, and stable robot control, the system can achieve a welding point positioning accuracy of ±0.03mm (3σ) and a single-point processing cycle of ≤85ms under complex working conditions such as illumination fluctuations (200-1500 lux) and splash interference. Compared with the traditional single algorithm solution, the recognition rate is improved by 23.6%, effectively supporting 0.1mm-level precision welding operations and greatly improving the quality and efficiency of steel cage welding.

Claims

1. An automated welding method for steel cage welding robots, characterized in that, Includes the following steps: Step S1: The main reinforcing bars are automatically fed by the storage mechanism, and the steel bars are cut to length and threaded through the roller conveyor system. The main reinforcing bars are positioned, aligned and clamped by the transmission device. Step S2: After each main rib is delivered, the automatic rotating mechanism of the skeleton drives the servo motor to rotate the internal long shaft by a preset angle, so that the stirrup support device reaches the next material placement station. Step S3: After the main reinforcement and stirrups are laid out, capture the image of the welding area, generate a probability heat map of the welding point, and obtain the three-dimensional coordinates of the welding point through SE(3) spatial coordinate transformation. In step S4, the six-axis robot, which is inverted and mounted on the cantilever structure, plans a collision-free path in the three-dimensional work space according to the spatial distribution of the welding task, and generates a smooth welding trajectory through adaptive B-spline interpolation. During the trajectory execution, a friction compensation model is introduced to suppress nonlinear friction interference and complete the skeleton welding. In step S3, generating the probability heatmap of the welding points includes the following sub-steps: Step S301: Preprocess the original image of the welding area to obtain a 256-level edge intensity map and semantic features extracted by the improved ResNet-50. Step S302: The preprocessed 256-level edge intensity map and the semantic features extracted by the improved ResNet-50 are input into the feature pyramid network for channel concatenation. Step S303: Cross-modal feature interaction is achieved by using a 3×3 convolution kernel on the concatenated features to generate a fused feature map of spatial and semantic domains. Step S304: The fused feature map is upsampled to the original image size using a bilinear interpolation algorithm to obtain a welding point probability heatmap with sub-pixel accuracy. In step S4, the improved RRT algorithm includes: Step S401: Define a three-dimensional search domain containing dynamic obstacles in the robot's workspace; Step S402: Node expansion is performed using an adaptive step size factor α, with an initial value of 1% to 3% of the maximum path length; and based on the Euclidean distance... Local space complexity and regulatory factors Dynamic correction α: ； Step S403: Extend the determination of the interference between the path segment and the steel cage structure by using the bounding box-triangular mesh cross detection method, and introduce a dynamic obstacle time series prediction model to evaluate the overlapping area of ​​the motion trajectory.

2. The automatic welding method for a rebar cage welding robot according to claim 1, characterized in that, In step S301, the preprocessing process includes: Step S3011: An edge feature map is generated using the Canny edge detection algorithm with adaptive gamma correction, wherein the function for adaptive gamma correction is: ， in: , , To calculate The mean of all pixels within a 5x5 window centered on the center; Step S3012: The edge feature map is processed based on the CLAHE algorithm. An adaptive local histogram equalization method with a fixed block size of 32×32 is adopted, and a brightness-sensitive limiting mechanism is introduced to generate a 256-level edge intensity map. The limiting threshold is: ， in: In pixels The mean of the local block centered on it; The mean of the entire image. The maximum amplitude limit threshold; Step S3013: Extract semantic features of the original image using the improved ResNet-50.

3. The automatic welding method for a rebar cage welding robot according to claim 2, characterized in that, In step S3013, the process of extracting semantic features using the improved ResNet-50 is as follows: In the third and fourth residual blocks of ResNet-50, deformable convolutional layers are embedded after the 3×3 convolutional layers in each residual block to adapt to the deformation characteristics of the weld points. After the residual block output, a coordinate attention module is cascaded to perform global average pooling on the width and height directions of the original image to enhance the geometric features of the welding points.

4. The automatic welding method for a rebar cage welding robot according to claim 3, characterized in that, The steps of global average pooling specifically include: Step S30131, Multi-dimensional spatial pooling: For the feature map of each channel, average it in the width and height directions respectively to obtain the shape as follows: , tensor; Step S30132, cross-channel weight generation; Generate response weights for specific spatial dimensions respectively: ， ， Here, CN is a 1×1 convolution used to integrate information across channels and generate weights. The ReLU activation function introduces nonlinearity. Step S30133, Fusion Modulation: First and Broadcast to input features Same spatial dimensions The comprehensive weight map is obtained. ; Input features again Element-wise multiplication with the weighted graph: ， The edges of the weld points are enhanced, while the background area is suppressed; Step S30134, residual connection output: Preserve original input The residual path, the final output is: 。 5. The automatic welding method for a rebar cage welding robot according to claim 1, characterized in that, In step S3, the SE(3) spatial coordinate transformation specifically includes the following steps: Step S305, establish the closed-loop transformation matrix from the image coordinate system to the tool coordinate system based on Eye-in-Hand calibration, specifically including the following sub-steps: Step S3051: Collect images and pose data from at least 6 different viewpoints, and use the DLT algorithm to solve the homography matrix to construct a mapping model; Step S3052: The initial values ​​of the camera extrinsic parameters are decomposed using the Tsai-Lenz radial constraint method, and the rotation matrix R and translation vector t are iteratively optimized using the Levenberg-Marquardt algorithm to minimize the projection error function. ， in, To calibrate the world coordinates of the board corner points, For image coordinates, The projection function of the camera intrinsic parameters, up to the change in error. ; Step S3053: Combining the pseudo-inverse solution strategy, the optimized camera pose is... With the robot's end-effector pose Substitute into the closed-loop transformation matrix: ， Expand into a system of linear equations For the coefficient matrix After applying Tikhonov regularization, the pseudoinverse is solved by singular value decomposition. The least squares solution is obtained: ； Step S3054 verifies consistency through repeated random posture tests to ensure that the three-dimensional coordinate transformation accuracy of the welding point meets the sub-millimeter positioning requirements.

6. The automatic welding method for a rebar cage welding robot according to claim 5, characterized in that, In step S4, the adaptive B-spline interpolation process dynamically sets the control point density and node vector spacing based on the spatial angle and curvature between welding points to generate a value that satisfies the speed requirement. acceleration Constrained continuous trajectory.

7. The automatic welding method for a rebar cage welding robot according to claim 1, characterized in that, In step S4, the friction compensation model is as follows: ， in, This is the theoretical torque command; Joint angular velocity; For symbolic functions, and These are the viscous friction coefficient and the Coulomb friction coefficient identified offline, respectively.

8. The automatic welding method for a rebar cage welding robot according to claim 1, characterized in that, In step S2, the major axis rotates by a preset angle: ， Where N is the total number of main reinforcing bars around the circumference.

Images