Automatic welding method for composite woven dense pipe based on AI and image sensor

Abstract

The invention provides a composite woven dense pipe automatic welding method based on AI and an image sensor, which relates to the technical field of welding, and comprises the following steps: performing three-dimensional reconstruction on a dense pipe by adopting a binocular stereoscopic vision system to obtain point cloud data, generating a space contour model through point cloud registration and a grid reconstruction algorithm, and establishing a welding coordinate system; meanwhile, a molten pool image is obtained in real time through a laser triangulation sensor, and the molten pool temperature is dynamically adjusted based on a self-adaptive temperature field control algorithm. The welding precision and quality can be improved, the environmental interference influence is reduced, and efficient and automatic welding of the composite woven dense pipe is achieved.

Description

Technical Field

[0001] This invention relates to the field of welding technology, and in particular to an automated welding method for composite braided tubes based on AI and image sensors. Background Technology

[0002] Composite braided tubing is widely used in high-end manufacturing fields such as aerospace, petrochemicals, and nuclear power equipment due to its excellent flexibility, corrosion resistance, and sealing performance. These tubings are typically made of multiple layers of metal or alloy wires, resulting in a complex structure and non-uniform surface, posing significant challenges to welding. Traditional welding of composite braided tubing relies heavily on manual operation, and the welding quality is often limited by the operator's experience and skill level, making it difficult to meet the requirements for high-precision, high-quality welding.

[0003] With the increasing level of industrial automation, vision-sensing-based automated welding technology is gradually being applied to the processing of composite braided tubing. Existing technologies typically employ monocular vision systems or binocular vision systems with fixed parameters to acquire workpiece information and perform welding operations using preset welding parameters. However, these technologies still have significant shortcomings in practical applications.

[0004] Existing vision systems are susceptible to changes in lighting and lens distortion in complex environments, resulting in insufficient accuracy of the acquired 3D data. This makes it impossible to accurately capture the subtle features of the composite braided tube surface, thus affecting the accuracy of subsequent welding trajectory planning. Secondly, existing welding trajectory planning methods are mostly based on simplified models or predefined parameters, which are difficult to adapt to the complex spatial structure and surface feature variations of composite braided tubes, leading to deviations between the welding trajectory and actual requirements. Summary of the Invention

[0005] This invention provides an automated welding method for composite braided tubing based on AI and image sensors, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides an automated welding method for composite braided tubing based on AI and image sensors, comprising: A binocular stereo vision system is used to perform three-dimensional reconstruction of the composite braided tube. The binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of the surface features of the composite braided tube, thereby obtaining three-dimensional point cloud reconstruction data. Based on the three-dimensional point cloud reconstruction data, a spatial contour model of the composite braided tube is generated through point cloud registration and mesh reconstruction algorithms. A welding coordinate system is established based on the spatial contour model. The weld trajectory point sequence is calculated based on the welding coordinate system to generate a welding trajectory planning scheme. After welding is completed, the target composite braided tube is selectively cut off. During the welding process, the molten pool image is acquired in real time by a laser triangulation sensor. The boundary contour and surface features of the molten pool are extracted based on the image segmentation algorithm. An adaptive temperature field control algorithm is used to dynamically adjust the temperature of the molten pool. The adaptive temperature field control algorithm calculates the temperature field distribution of the molten pool based on the grayscale distribution characteristics of the molten pool image. Based on the temperature field distribution data, the boundary contour and temperature parameters of the molten pool are determined. Based on the temperature parameters and the corresponding control gain coefficient, the current parameters and voltage parameters are determined.

[0007] The binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of surface features of the composite braided tube, resulting in 3D point cloud reconstruction data including: The observation image is acquired through a binocular stereo vision system. The observation image is adaptively compensated based on the scene reflectivity and illumination components. The observation image is decomposed into multiple scale levels, and a Gaussian filter is applied to each scale level. The compensated image is obtained through weighted fusion. Establish a mapping relationship between pixel coordinates and physical coordinates of the compensated image, calculate the radial distortion coefficient and tangential distortion coefficient of the compensated image based on the mapping relationship, correct the image distortion based on the radial distortion coefficient and tangential distortion coefficient, and obtain the corrected image. Based on the distortion correction result of the corrected image, a binary coding sequence that meets the spatial coding requirements is generated. A grayscale coded structured light projection pattern with a unique correspondence is constructed according to the binary coding sequence. The grayscale coded structured light projection pattern is projected onto the corrected image. The structured light stripe features are extracted by multi-phase offset calculation. The structured light stripe features include the stripe center line position and stripe grayscale distribution information. Based on the fringe centerline position and fringe grayscale distribution information of the structured light fringe features, a one-to-one correspondence between image coordinate points and spatial position points is established. Based on the one-to-one correspondence, the camera focal length parameters, binocular baseline length, and pixel disparity values ​​are calculated to generate three-dimensional point cloud reconstruction data.

[0008] Based on the 3D point cloud reconstruction data, a spatial contour model of the composite braided dense tube is generated through point cloud registration and mesh reconstruction algorithms. A welding coordinate system is then established based on this spatial contour model. The weld trajectory point sequence is calculated using the welding coordinate system, and a welding trajectory planning scheme is generated, including: The three-dimensional point cloud reconstruction data is reconstructed using Poisson reconstruction to generate an initial mesh. The neighborhood topology of the vertices of the initial mesh is calculated. Based on the neighborhood topology, the Laplace weight coefficient is determined. The vertices of the initial mesh are iteratively smoothed according to the Laplace weight coefficient to generate an optimized composite braided tube surface mesh. Extract vertex coordinates from the optimized surface mesh, construct a point cloud covariance matrix, perform eigenvalue decomposition on the point cloud covariance matrix, take the eigenvector corresponding to the largest eigenvalue as the axial direction of the dense tube, and establish a welding coordinate system by combining the axial direction of the dense tube and the optimized surface mesh. In the welding coordinate system, each vertex in the surface mesh is taken as a candidate growth point. The Gaussian curvature value of each candidate growth point and the angular deviation between adjacent normal vectors are calculated to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. The local curvature values ​​of discrete sampling points are calculated on the weld feature lines to generate an initial control point sequence. A B-spline curve with parameter continuity is constructed using the control point sequence. Trajectory points are sampled on the B-spline curve according to a preset spatial step size and curvature constraint. The tool posture angle is calculated based on the spatial position of the trajectory points to generate a welding trajectory planning scheme that satisfies kinematic constraints. The welding trajectory planning scheme includes a trajectory point position sequence and a tool posture sequence.

[0009] In the welding coordinate system, each vertex in the surface mesh is used as a candidate growth point. The Gaussian curvature value and the angular deviation between adjacent normal vectors are calculated for each candidate growth point to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. The local curvature values ​​of discrete sampling points are calculated on the weld feature lines to generate an initial control point sequence, including: Each vertex in the surface mesh is taken as a candidate growth point. The Gaussian curvature value and the angle deviation between adjacent normal vectors of each candidate growth point are calculated. The Gaussian curvature value and the angle deviation are normalized and then weighted and summed to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. Calculate the local curvature value of discrete sampling points on the weld feature line, determine the curvature change trend between adjacent discrete sampling points, and determine the sampling points whose local curvature value is greater than the preset curvature threshold and whose curvature sign changes as feature control points to generate an initial control point sequence. Positional constraints and curvature continuity constraints are applied to the initial control point sequence. By minimizing the control point position deviation and the second-order difference curvature, the optimized control point sequence is obtained.

[0010] The adaptive temperature field control algorithm calculates the molten pool temperature field distribution based on the grayscale distribution characteristics of the molten pool image, determines the molten pool boundary contour and temperature parameters based on the molten pool temperature field distribution data, and determines the current and voltage parameters based on the temperature parameters and the corresponding control gain coefficient, including: The image of the molten pool is subjected to grayscale normalization to obtain normalized grayscale values. Based on the law of radiation, a mapping relationship between the normalized grayscale values ​​and temperature is established. Combined with ambient temperature parameters and radiation coefficient parameters, the temperature field distribution data of the molten pool is calculated. Based on the molten pool temperature field distribution data, the molten pool boundary profile is determined, and the maximum distance of the molten pool boundary profile in the horizontal and vertical directions is calculated to obtain the molten pool width parameter and the molten pool length parameter. At the same time, the maximum temperature value is extracted from the molten pool temperature field distribution data to obtain the molten pool highest temperature parameter, and the spatial gradient of the molten pool temperature field distribution data is calculated to obtain the temperature gradient parameter. Determine the temperature deviation between the highest temperature parameter of the molten pool and the preset target temperature, the gradient deviation between the temperature gradient parameter and the preset target gradient, and the width deviation between the molten pool width parameter and the preset target width; The optimized current parameters are obtained by multiplying the temperature deviation value and the gradient deviation value by the corresponding temperature control gain coefficient, normalizing them, and then superimposing them with the reference current. The optimized voltage parameters are obtained by multiplying the width deviation value by the width control gain coefficient, normalizing them, and then superimposing them with the reference voltage.

[0011] The method further includes: Based on the analysis results of the molten pool image, the real-time offset of the weld position is obtained. The offset is converted to the welding coordinate system, and the welding trajectory planning scheme is corrected in real time to obtain the compensated welding trajectory. The welding torch is then controlled to complete the welding operation along the compensated welding trajectory.

[0012] A second aspect of the present invention provides an automated welding system for composite braided tubing based on AI and image sensors, comprising: The first unit is used to perform three-dimensional reconstruction of the composite braided tube using a binocular stereo vision system. The binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of the surface features of the composite braided tube, thereby obtaining three-dimensional point cloud reconstruction data. The second unit is used to generate a spatial contour model of the composite braided tube based on the three-dimensional point cloud reconstruction data through point cloud registration and mesh reconstruction algorithms, establish a welding coordinate system based on the spatial contour model, calculate the weld trajectory point sequence based on the welding coordinate system, generate a welding trajectory planning scheme, and cut the target composite braided tube after selective welding is completed. The third unit is used to acquire molten pool images in real time during the welding process using a laser triangulation sensor, extract the molten pool boundary contour and surface features based on an image segmentation algorithm, and dynamically adjust the molten pool temperature using an adaptive temperature field control algorithm. The adaptive temperature field control algorithm calculates the molten pool temperature field distribution based on the grayscale distribution characteristics of the molten pool image, determines the molten pool boundary contour and temperature parameters based on the molten pool temperature field distribution data, and determines the current and voltage parameters based on the temperature parameters and the corresponding control gain coefficient.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: This invention uses a binocular stereo vision system combined with structured light projection technology for three-dimensional reconstruction, achieving high-precision acquisition of surface features of composite braided tubes. It effectively overcomes the effects of ambient light interference and lens distortion, and improves the accuracy of pre-welding preparation.

[0016] This invention establishes an accurate spatial contour model and welding coordinate system based on 3D point cloud reconstruction data, which can automatically generate optimized welding trajectory planning schemes, reduce errors in manual planning, and significantly improve the accuracy of welding paths and work efficiency.

[0017] This invention acquires molten pool information in real time through a laser triangulation sensor during the welding process and dynamically adjusts the molten pool temperature using an adaptive temperature field control algorithm, thereby realizing intelligent adjustment of welding parameters and effectively improving welding quality and stability. It is particularly suitable for automated precision welding of complex-shaped composite braided tubes. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of an automatic welding method for composite braided tubes based on AI and image sensors, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the automated welding method for composite braided tubing based on AI and image sensors according to an embodiment of the present invention. Figure 1 As shown, the method includes: A binocular stereo vision system is used to perform three-dimensional reconstruction of the composite braided tube. The binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of the surface features of the composite braided tube, thereby obtaining three-dimensional point cloud reconstruction data. Based on the three-dimensional point cloud reconstruction data, a spatial contour model of the composite braided tube is generated through point cloud registration and mesh reconstruction algorithms. A welding coordinate system is established based on the spatial contour model. The weld trajectory point sequence is calculated based on the welding coordinate system to generate a welding trajectory planning scheme. After welding is completed, the target composite braided tube is selectively cut off. During the welding process, the molten pool image is acquired in real time by a laser triangulation sensor. The boundary contour and surface features of the molten pool are extracted based on the image segmentation algorithm. An adaptive temperature field control algorithm is used to dynamically adjust the temperature of the molten pool. The adaptive temperature field control algorithm calculates the temperature field distribution of the molten pool based on the grayscale distribution characteristics of the molten pool image. Based on the temperature field distribution data, the boundary contour and temperature parameters of the molten pool are determined. Based on the temperature parameters and the corresponding control gain coefficient, the current parameters and voltage parameters are determined.

[0022] In one optional implementation, the binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of the surface features of the composite braided tube, obtaining 3D point cloud reconstruction data including: The observation image is acquired through a binocular stereo vision system. The observation image is adaptively compensated based on the scene reflectivity and illumination components. The observation image is decomposed into multiple scale levels, and a Gaussian filter is applied to each scale level. The compensated image is obtained through weighted fusion. Establish a mapping relationship between pixel coordinates and physical coordinates of the compensated image, calculate the radial distortion coefficient and tangential distortion coefficient of the compensated image based on the mapping relationship, correct the image distortion based on the radial distortion coefficient and tangential distortion coefficient, and obtain the corrected image. Based on the distortion correction result of the corrected image, a binary coding sequence that meets the spatial coding requirements is generated. A grayscale coded structured light projection pattern with a unique correspondence is constructed according to the binary coding sequence. The grayscale coded structured light projection pattern is projected onto the corrected image. The structured light stripe features are extracted by multi-phase offset calculation. The structured light stripe features include the stripe center line position and stripe grayscale distribution information. Based on the fringe centerline position and fringe grayscale distribution information of the structured light fringe features, a one-to-one correspondence between image coordinate points and spatial position points is established. Based on the one-to-one correspondence, the camera focal length parameters, binocular baseline length, and pixel disparity values ​​are calculated to generate three-dimensional point cloud reconstruction data.

[0023] This invention provides a method for real-time compensation of ambient light interference and lens distortion in a binocular stereo vision system through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of surface features of a composite braided tube, thereby obtaining three-dimensional point cloud reconstruction data.

[0024] The binocular stereo vision system first acquires the observation image. This system consists of two industrial cameras with a pixel resolution of 1920×1080, a lens focal length of 8mm, and a mounting distance of 200mm. After image acquisition, the system adaptively compensates for the observation image based on scene reflectivity and illumination components. Specifically, the system decomposes the observation image into five different scale levels, ranging from 0.5 to 2.0, with an interval of 0.3. A Gaussian filter is applied to each scale level, with a filter kernel size of 5×5 pixels and a standard deviation of 1.5. After filtering, the system calculates weight coefficients based on the gradient information of each scale level: 0.3 for areas with strong illumination, 0.5 for areas with medium brightness, and 0.2 for dark areas. These weight coefficients are then used to weight and fuse the images at each scale level to obtain the compensated image. Test results show that this compensation method can improve image brightness uniformity by approximately 37% and enhance detail information in dark areas by approximately 45%.

[0025] After obtaining the compensated image, the system establishes a mapping relationship between pixel coordinates and physical coordinates. The system uses a 10×8 checkerboard calibration board, with each square measuring 15mm×15mm. By acquiring images of the calibration board at different angles (at least 20 sets), the system extracts the corner coordinates of the checkerboard and establishes the correspondence between pixel coordinates and physical coordinates. Based on these corresponding points, the system uses iterative least squares to calculate the radial distortion coefficients k1=-0.2536, k2=0.1072, k3=-0.0185, and the tangential distortion coefficients p1=0.0023, p2=-0.0011. Based on these distortion coefficients, the system corrects the image. The correction algorithm maps the original pixel positions to the corrected positions, achieving a correction accuracy of ±0.1 pixels. After correction, the linear distortion rate of the image is reduced from the original 6.8% to below 0.3%.

[0026] Based on the corrected image, the system generates a binary coded sequence that meets the spatial coding requirements. The coded sequence is 1024 bits long and uses an improved Gray code encoding method, ensuring that adjacent codes differ by only one bit, thus improving anti-interference capability. The system constructs a grayscale coded structured light projection pattern based on the binary coded sequence, with 256 grayscale levels and a projection resolution of 1280×800. The structured light projection is performed by a projector with a brightness of 3500 lumens, and the projection distance is set to 500mm to 800mm. The projected pattern forms stripes on the surface of the composite braided tube, with a stripe width of 2mm and an adjacent stripe spacing of 3mm. The system extracts stripe features using an eight-step phase-shifting method with a phase-shifting step size of π / 4, including sub-pixel accuracy of the stripe centerline position and stripe grayscale distribution information. The extracted stripe centerline accuracy can reach ±0.05 pixels, and the stripe grayscale recognition accuracy exceeds 98%.

[0027] The system establishes a one-to-one correspondence between image coordinate points and spatial location points based on stripe features. For each pair of corresponding points captured by the binocular camera, the system calculates the pixel disparity value using a block matching algorithm with a matching window size of 11×11 pixels. Based on the principle of triangulation, the system calculates the 3D coordinates of spatial points using the calibrated camera focal length parameters (fx=1253.6 pixels, fy=1251.8 pixels) and the binocular baseline length (200 mm). For a typical composite braided tube (50 mm in diameter, 24 strands / inch), the system reconstructs a 3D point cloud data with a point density of 5 points / mm. 2 The spatial measurement accuracy is within ±0.1mm. The system also performs noise filtering on the point cloud data, removing outliers that deviate from the average depth by more than 2 standard deviations, accounting for approximately 3% of the total points. The final generated 3D point cloud reconstruction data contains detailed morphological features of the pipe surface, accurately reflecting the depth variations (approximately 0.5-1.2mm) and local geometric features of the woven texture.

[0028] Through the above-mentioned technologies, the binocular stereo vision system of the present invention can perform high-precision three-dimensional reconstruction of the surface of composite braided tubes under varying ambient light conditions (50-1000 lux), meeting the requirements of industrial inspection for accuracy and real-time performance, with a processing speed of up to 15 frames / second.

[0029] In one optional implementation, based on the 3D point cloud reconstruction data, a spatial contour model of the composite braided dense tube is generated through point cloud registration and mesh reconstruction algorithms. A welding coordinate system is then established based on the spatial contour model. A weld trajectory point sequence is calculated based on the welding coordinate system, and a welding trajectory planning scheme is generated, including: The three-dimensional point cloud reconstruction data is reconstructed using Poisson reconstruction to generate an initial mesh. The neighborhood topology of the vertices of the initial mesh is calculated. Based on the neighborhood topology, the Laplace weight coefficient is determined. The vertices of the initial mesh are iteratively smoothed according to the Laplace weight coefficient to generate an optimized composite braided tube surface mesh. Extract vertex coordinates from the optimized surface mesh, construct a point cloud covariance matrix, perform eigenvalue decomposition on the point cloud covariance matrix, take the eigenvector corresponding to the largest eigenvalue as the axial direction of the dense tube, and establish a welding coordinate system by combining the axial direction of the dense tube and the optimized surface mesh. In the welding coordinate system, each vertex in the surface mesh is taken as a candidate growth point. The Gaussian curvature value of each candidate growth point and the angular deviation between adjacent normal vectors are calculated to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. The local curvature values ​​of discrete sampling points are calculated on the weld feature lines to generate an initial control point sequence. A B-spline curve with parameter continuity is constructed using the control point sequence. Trajectory points are sampled on the B-spline curve according to a preset spatial step size and curvature constraint. The tool posture angle is calculated based on the spatial position of the trajectory points to generate a welding trajectory planning scheme that satisfies kinematic constraints. The welding trajectory planning scheme includes a trajectory point position sequence and a tool posture sequence.

[0030] After acquiring the 3D point cloud reconstruction data, Poisson reconstruction is performed on the data to generate an initial mesh model. The Poisson reconstruction algorithm transforms the disordered point cloud into an ordered mesh surface by solving the Poisson equation. The generated initial mesh contains the geometric features of a composite braided tube. For example, for braided tube point cloud data with a diameter of 30mm and a length of 150mm (containing approximately 15,000 points), Poisson reconstruction can yield an initial mesh containing approximately 8,000 faces and 4,000 vertices.

[0031] Calculate the neighborhood topology of the initial grid vertices. By analyzing the connections between each vertex and its neighboring vertices, establish a topology matrix between vertices. For each vertex vi in ​​the grid, find all vertices vj that are directly connected to it, forming the first-order neighborhood set N(i) of vertex vi. For example, for a typical internal grid vertex, its first-order neighborhood usually contains 6-8 neighboring vertices.

[0032] Based on neighborhood topological relationships, the Laplace weight coefficients are determined. A cotangent weighting scheme is used to calculate the weight value between each pair of adjacent vertices; the weight value is related to the geometric relationship between the vertices. For each neighboring vertex vj of vertex vi, the corresponding weight wij is calculated. The weight value depends on the spatial distance between vertices vi and vj and their surrounding angular relationships. For example, for a pair of adjacent vertices with a distance of 2mm, if the angle they form is small (e.g., 30 degrees), the weight value is large (e.g., 0.8); if the angle is large (e.g., 120 degrees), the weight value is small (e.g., 0.2).

[0033] The initial mesh vertices are iteratively smoothed using Laplacian weighting coefficients. In each iteration, the new position of each vertex is determined by the combination of its own position and the positions of its weighted neighboring vertices. Typically, 5-10 iterations are performed until the mesh surface smoothness reaches a preset threshold. For example, iteration stops when the average vertex displacement between two consecutive iterations is less than 0.01 mm. After iterative smoothing, an optimized composite braided dense-tube surface mesh is obtained, with significantly reduced surface noise and improved surface continuity.

[0034] Vertex coordinates are extracted from the optimized surface mesh to construct the point cloud covariance matrix. The set of all vertex coordinates is denoted as {p1, p2, ..., pn}. The center point p̄ of the point cloud is calculated, and then the covariance matrix C is calculated. The covariance matrix C is 3×3 in size, reflecting the distribution of the point cloud in various directions in three-dimensional space.

[0035] Eigenvalue decomposition is performed on the point cloud covariance matrix to obtain three eigenvalues ​​λ1≥λ2≥λ3 and corresponding eigenvectors v1, v2, v3. The eigenvector v1 corresponding to the largest eigenvalue λ1 represents the principal direction of the point cloud distribution, i.e., the axial direction of the dense tube. For example, for an approximately cylindrical dense tube, the largest eigenvalue is usually much larger than the other two eigenvalues ​​(e.g., λ1=120.5, λ2=4.8, λ3=4.2), and the corresponding eigenvector v1=(0.02, 0.06, 0.998) represents the direction of the central axis of the dense tube.

[0036] A welding coordinate system is established by combining the axial direction of the dense tube and the optimized surface mesh. The axial direction v1 of the dense tube is taken as the Z-axis of the welding coordinate system, and the direction perpendicular to the Z-axis is chosen as the X-axis. The Y-axis direction is calculated using the cross product to construct an orthogonal coordinate system. The origin of the coordinate system can be set at the centroid of the mesh. This coordinate system serves as a reference system for subsequent welding trajectory planning.

[0037] In the welding coordinate system, each vertex in the surface mesh is considered as a candidate growth point. The Gaussian curvature value and the deviation of the angle between adjacent normal vectors are calculated for each candidate point. Gaussian curvature represents the inherent curvature characteristics at a point, and the deviation of the angle between normal vectors reflects the degree of change in the local surface. For example, for vertices at the weld location, the Gaussian curvature is usually large (e.g., 0.015 mm^-2), and the deviation of the angle between adjacent normal vectors is also large (e.g., the average deviation is greater than 15 degrees).

[0038] Based on the calculated Gaussian curvature and the deviation of the normal vector angle, an initial seed point (e.g., a point with a curvature value greater than 0.01 mm^-2) is selected from the high curvature region to determine the expansion direction of the region growth. Using a region growth algorithm, starting from the seed point, the region expands towards adjacent vertices, selecting adjacent points that meet certain conditions (e.g., a deviation of the normal vector angle greater than 12 degrees) to be added to the growth region, forming continuous weld feature lines.

[0039] Calculate the local curvature values ​​of discrete sampling points along the weld feature line, and select points with significant curvature changes (e.g., curvature change rate greater than 0.002 mm^-1) as control points to generate an initial control point sequence. For example, for a weld feature line approximately 200 mm long, 15-25 control points can be selected. The spacing between control points is denser in areas with large curvature changes (e.g., 5-8 mm) and sparser in areas with gentle curvature changes (e.g., 10-15 mm).

[0040] A cubic B-spline curve with parametric continuity is constructed using a sequence of control points. The B-spline curve guarantees second-order parametric continuity, ensuring a smooth transition of the welding trajectory. The node vectors and basis functions of the B-spline are set, and the points on the curve are calculated. For example, for a sequence containing 20 control points, the generated B-spline curve can be represented by 100-200 points, ensuring that the curve accurately describes the weld shape.

[0041] On the B-spline curve, sample trajectory points according to a preset spatial step size (e.g., 1-2 mm) and curvature constraints (e.g., step size reduced by 50% in areas with large curvature changes). Based on the spatial position of the trajectory points and the local tangent vector, normal vector, and binormal vector of the curve, calculate the tool attitude angles, including the tool's pitch angle, yaw angle, and roll angle. For example, the welding torch typically maintains a working posture with an angle of 15-20 degrees to the normal vector of the weld surface to obtain the best welding effect.

[0042] The final result is a welding trajectory planning scheme that satisfies kinematic constraints, including a sequence of trajectory point positions and a corresponding tool posture sequence. For a typical braided tube weld, the trajectory planning scheme may contain 300-500 trajectory points, each containing position coordinates (x, y, z) and posture angles (α, β, γ), enabling precise welding of the composite braided tube.

[0043] In one optional implementation, in the welding coordinate system, each vertex in the surface mesh is used as a candidate growth point. The Gaussian curvature value and the angular deviation between adjacent normal vectors are calculated for each candidate growth point to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. The local curvature values ​​of discrete sampling points are calculated on the weld feature lines to generate an initial control point sequence, including: Each vertex in the surface mesh is taken as a candidate growth point. The Gaussian curvature value and the angle deviation between adjacent normal vectors of each candidate growth point are calculated. The Gaussian curvature value and the angle deviation are normalized and then weighted and summed to determine the expansion direction of the region growth. Feature points of the surface mesh are continuously extracted along the expansion direction to form weld feature lines. Calculate the local curvature value of discrete sampling points on the weld feature line, determine the curvature change trend between adjacent discrete sampling points, and determine the sampling points whose local curvature value is greater than the preset curvature threshold and whose curvature sign changes as feature control points to generate an initial control point sequence. Positional constraints and curvature continuity constraints are applied to the initial control point sequence. By minimizing the control point position deviation and the second-order difference curvature, the optimized control point sequence is obtained.

[0044] This invention proposes a method for extracting weld feature lines based on Gaussian curvature and the deviation of the angle between the normal vectors. This method identifies the weld position through the geometric characteristics of the surface mesh and generates an accurate sequence of control points.

[0045] During implementation, the surface mesh data of the workpiece to be welded is first acquired, and a welding coordinate system is established. The surface mesh is typically composed of vertices and patches, with each vertex possessing three-dimensional coordinate information. For each vertex in the surface mesh, its Gaussian curvature value is calculated. Gaussian curvature is an important indicator of surface geometry and can be obtained by multiplying the principal curvatures around the vertex. For example, for a vertex with principal curvatures k1=0.05 and k2=-0.2, the Gaussian curvature of that point is -0.01. Simultaneously, the angular deviation between the normal vectors of this vertex and its adjacent vertices is calculated. The angular deviation reflects the degree of surface curvature and is typically larger in the weld region. For example, if the normal vector of a vertex is [0.1, 0.2, 0.9], and the normal vectors of its adjacent vertices are [0.15, 0.25, 0.85], the angular deviation between them is approximately 0.063 radians.

[0046] To comprehensively consider the influence of Gaussian curvature and the deviation of the normal vector angle, these two indicators need to be normalized. Normalization maps the Gaussian curvature value range from [-0.5, 0.5] to [0, 1], and the normal vector angle deviation from [0, π / 2] to [0, 1]. After normalization, the two indicators are weighted and summed. The weights can be adjusted according to the actual application; for example, the weight for Gaussian curvature is 0.6, and the weight for the normal vector angle deviation is 0.4. The weighted sum value is called the feature significance, used to determine the expansion direction of the region growth.

[0047] The region growth process begins with the vertex with the highest feature saliency, for example, a point with a feature saliency value of 0.85. Neighboring vertices with feature saliency greater than a preset threshold (e.g., 0.7) are searched around this point, and the vertex with the highest feature saliency is selected as the next growth point. This process is repeated, continuously extracting feature points from the surface mesh along the direction of high feature saliency to form weld feature lines. To avoid the region growth getting trapped in local optima, the continuity of the current growth direction is considered each time the next growth point is selected, ensuring the smoothness of the weld feature lines.

[0048] After obtaining the weld feature line, discrete sampling points are obtained by uniformly sampling along the feature line. The sampling interval can be set to 1.5 times the average side length of the grid; for example, if the average side length of the grid is 2 mm, the sampling interval is 3 mm. For each sampling point, its local curvature value is calculated. The local curvature can be obtained by fitting a circle to three points near the sampling point; the reciprocal of the circle's radius is the curvature. The curvature change trend between adjacent sampling points is determined, paying particular attention to the locations where the curvature sign changes, as this usually indicates points of change in weld geometry.

[0049] Sampling points whose local curvature values ​​are greater than a preset curvature threshold and whose curvature signs change are identified as feature control points. For example, if the preset curvature threshold is 0.01, when a sampling point is detected with a curvature value of 0.015, while its adjacent sampling points have a curvature value of -0.012, this sampling point is marked as a feature control point. All feature control points identified in this way form the initial control point sequence.

[0050] The initial control point sequence may contain redundancy or be uneven, requiring optimization. The optimization process imposes two types of constraints: positional constraints and curvature continuity constraints. Positional constraints ensure that the optimized control points do not deviate too far from their original positions; the maximum allowed offset is 0.5 times the average side length of the original mesh, for example, 1 mm. Curvature continuity constraints ensure smooth curvature changes between adjacent control points, achieved by minimizing the second-order difference curvature of the control points.

[0051] The positions of the control points are adjusted using iterative optimization algorithms, such as gradient descent, until convergence conditions are met: the positional deviation is less than 0.1 mm, or the number of iterations reaches a maximum of 50. The optimized control point sequence is smoother while preserving the key geometric features of the weld.

[0052] To illustrate with a practical example, the surface mesh of a welded workpiece contains 5000 vertices. By calculating the eigenvalue of each vertex, the vertex with a eigenvalue of 0.92 is identified as the starting growth point. Region growth is performed along the direction of high eigenvalue, extracting a weld feature line containing 150 points. 50 sampling points are uniformly sampled along this feature line. After calculating the local curvature value of each sampling point, seven feature control points are identified, with curvature values ​​of 0.025, -0.018, 0.032, -0.027, 0.020, -0.022, and 0.015, respectively. After optimization, a smooth control point sequence is obtained, which can ultimately be used to guide the welding robot to accurately weld along the weld seam trajectory.

[0053] In one optional implementation, the adaptive temperature field control algorithm calculates the molten pool temperature field distribution based on the grayscale distribution characteristics of the molten pool image, determines the molten pool boundary contour and temperature parameters based on the molten pool temperature field distribution data, and determines the current and voltage parameters based on the temperature parameters and the corresponding control gain coefficient, including: The image of the molten pool is subjected to grayscale normalization to obtain normalized grayscale values. Based on the law of radiation, a mapping relationship between the normalized grayscale values ​​and temperature is established. Combined with ambient temperature parameters and radiation coefficient parameters, the temperature field distribution data of the molten pool is calculated. Based on the molten pool temperature field distribution data, the molten pool boundary profile is determined, and the maximum distance of the molten pool boundary profile in the horizontal and vertical directions is calculated to obtain the molten pool width parameter and the molten pool length parameter. At the same time, the maximum temperature value is extracted from the molten pool temperature field distribution data to obtain the molten pool highest temperature parameter, and the spatial gradient of the molten pool temperature field distribution data is calculated to obtain the temperature gradient parameter. Determine the temperature deviation between the highest temperature parameter of the molten pool and the preset target temperature, the gradient deviation between the temperature gradient parameter and the preset target gradient, and the width deviation between the molten pool width parameter and the preset target width; The optimized current parameters are obtained by multiplying the temperature deviation value and the gradient deviation value by the corresponding temperature control gain coefficient, normalizing them, and then superimposing them with the reference current. The optimized voltage parameters are obtained by multiplying the width deviation value by the width control gain coefficient, normalizing them, and then superimposing them with the reference voltage.

[0054] In this embodiment, the adaptive temperature field control algorithm includes four main steps: image preprocessing, temperature field calculation, molten pool feature extraction, and control parameter optimization. The image preprocessing step performs grayscale normalization on the acquired molten pool image. The temperature field calculation step calculates the molten pool temperature field distribution based on the normalized grayscale values. The molten pool feature extraction step determines the molten pool boundary contour and temperature parameters. The control parameter optimization step determines the optimized current and voltage parameters based on the deviation between the temperature parameters and the target values.

[0055] In the image preprocessing step, images of the molten pool during the welding process are acquired using a high-speed camera at a rate of 500 frames per second, with an image resolution of 1024×768 pixels. The acquired raw images typically contain various noises and interferences, requiring preprocessing to improve the accuracy of subsequent analysis. Preprocessing includes three operations: image cropping, median filtering, and grayscale normalization. Image cropping extracts the central 512×512 pixel region of the original image as the molten pool region of interest. Median filtering uses a 5×5 filter window to process the cropped image, effectively removing salt-and-pepper noise. Grayscale normalization maps the image's grayscale values ​​to between 0 and 1. The normalized grayscale value is calculated by subtracting the minimum grayscale value from the original grayscale value and then dividing by the difference between the maximum and minimum grayscale values.

[0056] In the temperature field calculation step, a mapping relationship between normalized grayscale values ​​and temperature is established based on the law of radiation. According to the Stephen-Boltzmann law, the radiation intensity of an object is proportional to the fourth power of its temperature. Considering the nonlinear response characteristics of the camera, this implementation adopts a modified radiation model, expressing the relationship between the normalized grayscale value G and the temperature T as: temperature T equals the sum of the ambient temperature and the nonlinear function of the normalized grayscale value. The ambient temperature parameter is set to 298K, and the radiation coefficient parameter is determined according to the material properties; for typical steel, this coefficient is taken as 0.8. Through this mapping relationship, the normalized grayscale value of each pixel in the molten pool image can be converted into the corresponding temperature value, thereby obtaining the temperature field distribution data of the entire molten pool area.

[0057] In practical applications, temperature field calculations need to consider the influence of camera parameters and environmental factors. To ensure calculation accuracy, thermocouples are used to measure actual temperatures at specific points to calibrate the temperature mapping model. During calibration, five high-temperature thermocouples are placed at different locations in the molten pool, and the actual temperature values ​​and corresponding image grayscale values ​​are recorded. Data points are obtained through multiple experiments, and the optimal mapping parameters are obtained by fitting using the least squares method. Experiments show that the average error between the temperature calculated by this method and the actual measured temperature is within 3%, meeting the control accuracy requirements.

[0058] In the molten pool feature extraction step, the molten pool boundary profile and key temperature parameters are determined based on the calculated temperature field distribution data. The molten pool boundary profile is determined using a temperature threshold method, identifying regions with temperatures above the material's melting point as molten pool areas. For commonly used low-carbon steel, the melting point temperature threshold is set to 1723K. By connecting points with temperatures equal to the threshold, the closed boundary profile of the molten pool is obtained. After obtaining the boundary profile, the maximum distance of the molten pool in the horizontal direction (X-axis) is calculated to obtain the molten pool width parameter, and the maximum distance in the vertical direction (Y-axis) is calculated to obtain the molten pool length parameter. Simultaneously, the maximum temperature value is extracted from the temperature field distribution data as the highest temperature parameter of the molten pool. The temperature gradient parameter is calculated using the central difference method, calculating the spatial derivatives of the temperature field distribution data in both the X and Y directions, and then calculating the gradient magnitude as the temperature gradient parameter.

[0059] In one specific example, the temperature in the central region of the processed molten pool image reached 2350K, the molten pool boundary temperature was 1723K, the molten pool width was 8.6mm, the molten pool length was 12.4mm, and the temperature gradient was 86K / mm. These parameters reflect the geometric characteristics and temperature distribution features of the molten pool under the current welding condition, and are important bases for subsequent control optimization.

[0060] In the control parameter optimization step, the optimized current and voltage parameters are calculated based on the deviation between the extracted temperature parameters and the preset target values, combined with the corresponding control gain coefficients. The preset target values ​​are determined according to the material properties and welding requirements. For a specific low-carbon steel welding task, the preset target temperature is 2200K, the preset target gradient is 80K / mm, and the preset target width is 8.0mm.

[0061] First, calculate the temperature deviation, gradient deviation, and width deviation. The temperature deviation is equal to the highest molten pool temperature parameter minus the preset target temperature; the gradient deviation is equal to the temperature gradient parameter minus the preset target gradient; and the width deviation is equal to the molten pool width parameter minus the preset target width. For the example above, the temperature deviation is 150K, the gradient deviation is 6K / mm, and the width deviation is 0.6mm.

[0062] Then, the control gain coefficients were determined. The temperature control gain coefficient Kt was set to 0.05 A / K, the gradient control gain coefficient Kg was set to 0.2 A / (K / mm), and the width control gain coefficient Kw was set to 2.0 V / mm. These coefficient values ​​were obtained through extensive experiments and optimization and can be fine-tuned according to specific welding tasks.

[0063] The optimization of current parameters considers both temperature deviation and gradient deviation. The temperature deviation value is multiplied by the temperature control gain coefficient Kt to obtain the temperature control increment, and the gradient deviation value is multiplied by the gradient control gain coefficient Kg to obtain the gradient control increment. The sum of these two increments is normalized by dividing the increment value by the maximum allowable increment value (set to 10A), and then superimposed with the reference current to obtain the optimized current parameters. The reference current is determined based on the welding material and thickness; for the low-carbon steel plate (5mm thick) in the example, the reference current is set to 180A.

[0064] Voltage parameter optimization primarily considers width deviation. The width deviation value is multiplied by the width control gain coefficient Kw to obtain the width control increment. After normalization (with the maximum allowable increment set to 5V), this increment is superimposed on the reference voltage to obtain the optimized voltage parameters. The reference voltage is set to 22V.

[0065] In the above example, multiplying the temperature deviation of 150K by the temperature control gain coefficient of 0.05A / K yields a temperature control increment of 7.5A. Multiplying the gradient deviation of 6K / mm by the gradient control gain coefficient of 0.2A / (K / mm) yields a gradient control increment of 1.2A. The sum of these two is 8.7A, which, after normalization, becomes 0.87, indicating an adjustment ratio of 87% to the reference current. Multiplying the width deviation of 0.6mm by the width control gain coefficient of 2.0V / mm yields a width control increment of 1.2V, which, after normalization, becomes 0.24, indicating an adjustment ratio of 24% to the reference voltage. The final optimized current parameter is 164.4A (reference current minus adjustment), and the voltage parameter is 23.2V (reference voltage plus adjustment).

[0066] The practical application of this adaptive control algorithm shows that, compared with traditional fixed-parameter welding, the fluctuation of the molten pool temperature is reduced by 65%, the consistency of the molten pool width is improved by 42%, and the weld quality is significantly improved. The system response time is less than 20 milliseconds, enabling real-time adjustment of welding parameters to adapt to various changes during the welding process. In 1000 consecutive welding tests, the molten pool temperature control accuracy remained within ±30K, and the molten pool width control accuracy remained within ±0.3mm, indicating that the algorithm has good stability and reliability.

[0067] In one optional implementation, the method further includes: Based on the analysis results of the molten pool image, the real-time offset of the weld position is obtained. The offset is converted to the welding coordinate system, and the welding trajectory planning scheme is corrected in real time to obtain the compensated welding trajectory. The welding torch is then controlled to complete the welding operation along the compensated welding trajectory.

[0068] In the process of real-time correction of the welding trajectory, a method based on molten pool image analysis is used to obtain the real-time offset of the weld position, and coordinate transformation and trajectory compensation are performed to ensure welding quality. This embodiment provides a complete method for real-time correction of the welding trajectory.

[0069] The real-time offset of the weld position is obtained by analyzing images of the molten pool acquired during the welding process. The molten pool images are acquired using a high-speed industrial camera mounted near the welding torch at a rate of 60 frames per second, with an image resolution of 1280×1024 pixels. After acquisition, the images undergo preprocessing, including grayscale conversion, Gaussian filtering for noise reduction, and image enhancement. The preprocessed images are then binarized with a threshold set to 180 to separate the molten pool area from the background. Morphological operations such as dilation and erosion are applied using a 5×5 structuring element to remove noise and small-area interference from the image.

[0070] The molten pool contour is extracted from the processed image, and an edge detection algorithm is used to identify the molten pool boundary and extract feature point sets. The molten pool contour is analyzed using a fitting algorithm to identify the weld position. Based on the difference between the ideal weld position and the actual detected weld position, the horizontal and vertical offsets are calculated. For example, when the actual weld position is detected to be offset by 2.3 mm horizontally and 1.5 mm vertically relative to a preset trajectory, these offsets will be used as input for subsequent trajectory correction.

[0071] After obtaining the offset, it needs to be transformed from the image coordinate system to the welding coordinate system. The image coordinate system is a two-dimensional planar coordinate system, with the origin located at the upper left corner of the image, the x-axis pointing to the right, and the y-axis pointing downwards. The welding coordinate system, on the other hand, is a three-dimensional spatial coordinate system, with the origin located at the reference point of the welding worktable, and the x, y, and z axes corresponding to the three axes of the robot base coordinate system. During the coordinate transformation, the transformation relationship between the image coordinate system and the camera coordinate system is first obtained through camera calibration. Camera calibration uses a checkerboard calibration board with a size of 9×7 grids, each grid having a side length of 25 mm. Twenty sets of calibration images from different angles are acquired, and the intrinsic parameter matrix and distortion coefficients are calculated.

[0072] Subsequently, hand-eye calibration was used to determine the transformation relationship between the camera coordinate system and the robot coordinate system. Hand-eye calibration employed a dedicated calibration tool, collecting calibration point data under different poses, with at least eight sets of data collected to ensure calibration accuracy. The rotation matrix R and translation vector T were calculated to establish the transformation relationship between the two coordinate systems. After the coordinate transformation, the offset was mapped to the welding coordinate system, representing a positional deviation in three-dimensional space, which was used for subsequent trajectory correction.

[0073] The welding trajectory planning scheme is corrected in real time based on the transformed offset. The original welding trajectory is generated offline and contains a series of key points, each containing position coordinates and attitude information. The trajectory correction adopts a real-time compensation strategy, adding the detected offset as a compensation value to the original trajectory. For example, the original trajectory point P(x, y,z) becomes P'(x+dx, y+dy, z+dz) after compensation, where dx, dy, and dz are the transformed offset components.

[0074] To ensure smooth trajectory correction, a sliding window filtering method is used to process the offset, with the window size set to 5 sampling points to reduce the impact of abrupt changes on the trajectory. Simultaneously, an offset change rate limit is set, ensuring that the offset change between two adjacent corrections does not exceed 1.5 mm to prevent abrupt trajectory changes from affecting welding quality. For cumulative offset, a threshold limit of ±8 mm is set; exceeding this threshold triggers an anomaly handling mechanism to ensure system safety.

[0075] The compensated welding trajectory data is transmitted in real time to the robot control system via an industrial communication network, controlling the welding torch to move along the corrected trajectory. The control system employs a feedforward control strategy to predict trajectory changes in advance and reduce system response lag. The welding torch speed is set within the range of 5-10 mm / s according to welding process requirements to ensure welding quality. Simultaneously, the distance between the welding torch and the workpiece is monitored in real time and maintained within a range of 12±1 mm to adapt to different welding conditions.

[0076] In practical applications, this method was successfully applied to the butt welding process of 8 mm thick steel plates. The initial trajectory was set along a straight weld seam with a length of 500 mm. During welding, the weld seam experienced a maximum offset of 3.7 mm due to thermal deformation of the workpiece. By using this method to detect the offset in real time and correct the trajectory, the final weld seam position accuracy was controlled within ±0.5 mm, and the welding quality met the process requirements. The system response time is less than 100 milliseconds, effectively handling various interference factors during the welding process.

[0077] This method is applicable to various welding processes, including TIG welding, MIG welding, and laser welding, and is particularly suitable for automated welding scenarios requiring high precision. Through molten pool image analysis and real-time trajectory correction, it significantly improves welding quality and production efficiency.

[0078] This invention relates to an automated welding system for composite braided tubing based on AI and image sensors, comprising: The first unit is used to perform three-dimensional reconstruction of the composite braided tube using a binocular stereo vision system. The binocular stereo vision system compensates for ambient light interference and lens distortion in real time through a dynamic calibration algorithm, and combines structured light projection technology to determine the acquisition accuracy of the surface features of the composite braided tube, thereby obtaining three-dimensional point cloud reconstruction data. The second unit is used to generate a spatial contour model of the composite braided tube based on the three-dimensional point cloud reconstruction data through point cloud registration and mesh reconstruction algorithms, establish a welding coordinate system based on the spatial contour model, calculate the weld trajectory point sequence based on the welding coordinate system, generate a welding trajectory planning scheme, and cut the target composite braided tube after welding is completed. The third unit is used to acquire molten pool images in real time during the welding process using a laser triangulation sensor, extract the molten pool boundary contour and surface features based on an image segmentation algorithm, and dynamically adjust the molten pool temperature using an adaptive temperature field control algorithm. The adaptive temperature field control algorithm calculates the molten pool temperature field distribution based on the grayscale distribution characteristics of the molten pool image, determines the molten pool boundary contour and temperature parameters based on the molten pool temperature field distribution data, and determines the current and voltage parameters based on the temperature parameters and the corresponding control gain coefficient.

[0079] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0080] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0081] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI and image sensor based composite braided tight tube automatic welding method, characterized by, The method comprises the following steps: Three-dimensional reconstruction of the composite braided dense pipe is performed by using a binocular stereo vision system, wherein the binocular stereo vision system compensates for environmental light interference and lens distortion in real time through a dynamic calibration algorithm, determines the collection accuracy of the surface features of the composite braided dense pipe in combination with a structured light projection technology, and obtains three-dimensional point cloud reconstruction data; According to the three-dimensional point cloud reconstruction data, a spatial contour model of the composite braided dense pipe is generated through a point cloud registration and mesh reconstruction algorithm, a welding coordinate system is established based on the spatial contour model, a sequence of welding seam track points is calculated according to the welding coordinate system, and a welding track planning scheme is generated; and the target composite braided dense pipe is selectively cut off after welding is completed; During the welding process, a laser triangulation sensor is used to acquire molten pool images in real time, an image segmentation algorithm is used to extract the molten pool boundary contour and the molten pool surface features, and an adaptive temperature field control algorithm is used to dynamically adjust the molten pool temperature, wherein the adaptive temperature field control algorithm calculates the molten pool temperature field distribution according to the gray distribution characteristics of the molten pool image, determines the molten pool boundary contour and the temperature parameters based on the molten pool temperature field distribution data, and determines the current and voltage parameters based on the temperature parameters and the corresponding control gain coefficients.

2. The method of claim 1, wherein, The binocular stereo vision system compensates for environmental light interference and lens distortion in real time through a dynamic calibration algorithm, determines the collection accuracy of the surface features of the composite braided dense pipe in combination with a structured light projection technology, and obtains three-dimensional point cloud reconstruction data, which comprises the following steps: An observation image is acquired by using a binocular stereo vision system, the observation image is adaptively compensated based on scene reflectivity and light components, the observation image is decomposed into multiple scale levels, and a Gaussian filter is applied to each scale level, a compensated image is obtained through weighted fusion; A mapping relationship between pixel coordinates and physical coordinates of the compensated image is established, the radial distortion coefficient and the tangential distortion coefficient of the compensated image are calculated according to the mapping relationship, the image distortion is corrected based on the radial distortion coefficient and the tangential distortion coefficient, and a corrected image is acquired; Based on the distortion correction result of the corrected image, a binary coding sequence meeting the spatial coding requirement is generated, a gray coding structured light projection pattern having a unique corresponding relationship is constructed according to the binary coding sequence, the gray coding structured light projection pattern is projected onto the corrected image, a structured light fringe feature is extracted through multi-phase offset calculation, and the structured light fringe feature comprises a fringe center line position and fringe gray distribution information; According to the fringe center line position and the fringe gray distribution information of the structured light fringe feature, a one-to-one corresponding relationship between image coordinate points and spatial position points is established, the camera focal length parameter, the binocular baseline length and the pixel disparity value are calculated based on the one-to-one corresponding relationship, and three-dimensional point cloud reconstruction data is generated.

3. The method of claim 1, wherein, According to the three-dimensional point cloud reconstruction data, a spatial contour model of the composite braided dense pipe is generated through a point cloud registration and mesh reconstruction algorithm, a welding coordinate system is established based on the spatial contour model, a sequence of welding seam track points is calculated according to the welding coordinate system, and a welding track planning scheme is generated, which comprises the following steps: The three-dimensional point cloud reconstruction data is subjected to Poisson reconstruction to generate an initial mesh, a neighborhood topological relationship of vertices of the initial mesh is calculated, a Laplace weight coefficient is determined based on the neighborhood topological relationship, and the vertices of the initial mesh are iteratively smoothed according to the Laplace weight coefficient to generate a surface mesh of the optimized composite braided dense tube; The vertex coordinates in the optimized surface mesh are extracted, a point cloud covariance matrix is constructed, the point cloud covariance matrix is subjected to eigenvalue decomposition, a feature vector corresponding to a maximum eigenvalue is taken as a dense tube axial direction, and a welding coordinate system is established in combination with the dense tube axial direction and the optimized surface mesh; In the welding coordinate system, each vertex in the surface mesh is taken as a candidate growth point, a Gaussian curvature value and an angle deviation of an adjacent normal vector of each candidate growth point are calculated, an expansion direction of region growth is determined, feature points of the surface mesh are continuously extracted along the expansion direction, and a weld feature line is formed; local curvature values of discrete sampling points on the weld feature line are calculated to generate an initial control point sequence; A B-spline curve with parameter continuity is constructed by using the control point sequence, trajectory points are sampled on the B-spline curve according to a preset spatial step and a curvature constraint, tool posture angles are calculated according to spatial positions of the trajectory points, and a welding trajectory planning scheme satisfying kinematic constraints is generated, the welding trajectory planning scheme including a sequence of trajectory point positions and a sequence of tool postures.

4. The method of claim 3, wherein, In the welding coordinate system, each vertex in the surface mesh is taken as a candidate growth point, a Gaussian curvature value and an angle deviation of an adjacent normal vector of each candidate growth point are calculated, an expansion direction of region growth is determined, feature points of the surface mesh are continuously extracted along the expansion direction, and a weld feature line is formed; Calculating local curvature values of discrete sampling points on the weld feature line to generate an initial control point sequence includes: Each vertex in the surface mesh is taken as a candidate growth point, a Gaussian curvature value and an angle deviation of an adjacent normal vector of each candidate growth point are calculated, and the Gaussian curvature value and the angle deviation are normalized and weighted summed to determine an expansion direction of region growth, feature points of the surface mesh are continuously extracted along the expansion direction, and a weld feature line is formed; Local curvature values of discrete sampling points on the weld feature line are calculated, a curvature change trend between adjacent discrete sampling points is judged, a sampling point with a local curvature value greater than a preset curvature threshold and a curvature sign change is determined as a feature control point, and an initial control point sequence is generated; Position constraints and curvature continuity constraints are applied to the initial control point sequence, and an optimized control point sequence is obtained by minimizing control point position deviations and second-order differential curvatures.

5. The method of claim 1, wherein, The adaptive temperature field control algorithm calculates a molten pool temperature field distribution according to a gray scale distribution feature of a molten pool image, determines a molten pool boundary contour and a temperature parameter based on the molten pool temperature field distribution data, and determines a current parameter and a voltage parameter based on the temperature parameter and a corresponding control gain coefficient includes: The molten pool image is subjected to gray scale normalization processing to obtain a normalized gray scale value, a mapping relationship between the normalized gray scale value and temperature is established based on the law of radiation, and a molten pool temperature field distribution data is calculated in combination with an ambient temperature parameter and a radiation coefficient parameter; A molten pool boundary contour is determined based on the molten pool temperature field distribution data, a maximum distance of the molten pool boundary contour in a horizontal direction and a vertical direction is calculated to obtain a molten pool width parameter and a molten pool length parameter, a maximum temperature value in the molten pool temperature field distribution data is extracted to obtain a molten pool highest temperature parameter, and a spatial gradient of the molten pool temperature field distribution data is calculated to obtain a temperature gradient parameter; A temperature deviation value of the molten pool highest temperature parameter from a preset target temperature, a gradient deviation value of the temperature gradient parameter from a preset target gradient, and a width deviation value of the molten pool width parameter from a preset target width are determined; The temperature deviation value and the gradient deviation value are multiplied by corresponding temperature control gain coefficients respectively and normalized, and then superimposed with a reference current to obtain an optimized current parameter, and the width deviation value is multiplied by a width control gain coefficient and normalized, and then superimposed with a reference voltage to obtain an optimized voltage parameter.

6. The method of claim 1, wherein, The method further comprises: According to the analysis result of the molten pool image, a real-time offset amount of a weld seam position is obtained, the offset amount is converted to a welding coordinate system, a real-time correction is made to a welding track planning scheme, a compensated welding track is obtained, and a welding torch is controlled to complete a welding operation along the compensated welding track.

7. An image sensor based composite braided tight tube automatic welding system for implementing the method according to any one of claims 1-6, characterized in that, Comprise: A first unit is configured to perform three-dimensional reconstruction on the composite braided dense pipe by using a binocular stereo vision system, wherein the binocular stereo vision system compensates for environmental light interference and lens distortion in real time by using a dynamic calibration algorithm, determines the collection accuracy of the surface features of the composite braided dense pipe in combination with a structured light projection technology, and obtains three-dimensional point cloud reconstruction data; A second unit is configured to generate a spatial contour model of the composite braided dense pipe by using a point cloud registration and mesh reconstruction algorithm according to the three-dimensional point cloud reconstruction data, establish a welding coordinate system based on the spatial contour model, calculate a weld seam track point sequence according to the welding coordinate system, and generate a welding track planning scheme; and selectively cut off the target composite braided dense pipe after welding is completed; A third unit is configured to obtain a molten pool image in real time by using a laser triangulation sensor during welding, extract a molten pool boundary contour and molten pool surface features based on an image segmentation algorithm, and dynamically adjust the molten pool temperature by using an adaptive temperature field control algorithm, wherein the adaptive temperature field control algorithm calculates a molten pool temperature field distribution according to the gray scale distribution characteristics of the molten pool image, determines a molten pool boundary contour and temperature parameters based on the molten pool temperature field distribution data, and determines current and voltage parameters based on the temperature parameters and corresponding control gain coefficients.

8. An electronic device, comprising: Comprise: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images