Deep learning-based real-time optimization method and system for welding trajectory of industrial robot

Abstract

The application provides a deep learning-based industrial robot welding track real-time optimization method and system, relates to the technical field of industrial robots, and comprises the following steps: collecting three-dimensional point cloud data and generating a high-precision curved surface through multi-scale processing, constructing a weld feature descriptor, optimizing the mapping relationship between process parameters and welding postures based on a multi-objective reward function, converting the mapping relationship into an initial motion track, and dynamically optimizing the track by adopting a segmented adaptive optimization model and a variable structure filtering algorithm to realize adaptive adjustment of the welding process. The application improves the welding precision and quality of complex curved surfaces and enhances the adaptability and stability of robot welding.

Description

Technical Field

[0001] This invention relates to industrial robot technology, and more particularly to a method and system for real-time optimization of welding trajectories for industrial robots based on deep learning. Background Technology

[0002] With the increasing automation of manufacturing, industrial robots are being used more and more widely in the welding field. Industrial robot welding technology can significantly improve production efficiency, welding quality, and working environment safety, especially when handling complex curved surface welding tasks, where it has irreplaceable advantages. Traditional industrial robot welding trajectory planning mainly relies on manual teaching or offline programming, completing welding tasks through pre-set paths and process parameters. With the development of deep learning and computer vision technologies, adaptive welding trajectory planning methods based on sensors and intelligent algorithms have received widespread attention. These methods can adjust welding parameters and trajectories in real time according to the actual welding environment and workpiece characteristics, improving welding quality and adaptability.

[0003] Existing technologies struggle to accurately acquire high-precision surface features and weld information when processing complex curved surface welding. Traditional point cloud data processing methods perform poorly in noise reduction and feature extraction when dealing with complex geometries, resulting in insufficient accuracy in subsequent weld recognition and trajectory planning, thus failing to meet the requirements of high-quality welding.

[0004] Existing welding trajectory optimization methods lack a joint optimization mechanism for welding process parameters and robot posture. Most methods treat welding process parameter optimization and trajectory planning as two independent processes, ignoring the complex interaction between process parameters and welding posture, making it difficult to achieve smooth and efficient trajectory while ensuring welding quality.

[0005] Existing trajectory optimization algorithms generally employ a globally uniform optimization strategy, lacking the ability to adaptively optimize for different curvature regions and welding difficulties. This method cannot flexibly adjust the optimization strategy according to the different characteristics of the welding trajectory, especially when dealing with complex structures such as variable cross-sections and irregular joints, making it difficult to balance welding quality and motion efficiency, thus affecting the final welding effect. Summary of the Invention

[0006] This invention provides a method and system for real-time optimization of welding trajectories for industrial robots based on deep learning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a real-time optimization method for welding trajectories of industrial robots based on deep learning, comprising:

[0008] Three-dimensional point cloud data of complex curved surfaces are collected, and the three-dimensional point cloud data are denoised and feature-enhanced using a multi-scale decomposition algorithm. High-precision curved surface data is generated by combining the curvature-constrained mesh reconstruction method.

[0009] A dynamic feature space is established on the high-precision surface data, and the weld trajectory point set in the dynamic feature space is extracted to construct a weld feature descriptor that integrates geometric features and topological relationships.

[0010] Based on the weld feature descriptor, welding current, voltage, and wire feed speed are used as joint optimization objectives. The optimal combination of process parameters and welding posture sequence are calculated iteratively through a multi-objective reward function, and the mapping relationship between the optimal combination of process parameters and the welding posture sequence is established.

[0011] The mapping relationship is mapped to the robot joint space, and an improved fast collision detection algorithm is used for obstacle avoidance verification. The initial motion trajectory is generated in combination with dynamic constraints.

[0012] A segmented adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory.

[0013] The optimized initial motion trajectory is converted into robot control commands to achieve adaptive adjustment of the welding process.

[0014] The 3D point cloud data is denoised and feature-enhanced using a multi-scale decomposition algorithm, and high-precision surface data is generated by combining a curvature-constrained mesh reconstruction method.

[0015] The three-dimensional point cloud data is decomposed into multiple scales based on wavelet transform, and the local curvature change rate at each decomposition scale is calculated. The local curvature change rate is then used to perform multi-scale noise reduction on the three-dimensional point cloud data to obtain the noise-reduced point cloud data.

[0016] Calculate the local neighborhood covariance matrix for the denoised point cloud data, extract the eigenvalues ​​and eigenvectors of the local neighborhood covariance matrix, and construct a local geometric feature descriptor containing curvature information and normal vector information based on the eigenvalues ​​and eigenvectors.

[0017] The feature distance between the local geometric feature descriptors is calculated using the Gaussian kernel function. Based on the feature distance and the local point cloud density distribution, feature enhancement and density equalization are performed to generate feature-enhanced point cloud data.

[0018] A local curvature tensor field is constructed on the feature-enhanced point cloud data. The local curvature change is estimated based on the local curvature tensor field. The local curvature change is transformed into a curvature constraint condition. The sampling strategy is determined in combination with the point cloud density to generate a shape quality index that satisfies the local curvature change.

[0019] The shape quality index, the curvature constraint, and the point cloud fitting error are iteratively optimized and adjusted to adjust the mesh structure until convergence, at which point high-precision surface data is output.

[0020] A local curvature tensor field is constructed on the feature-enhanced point cloud data. Local curvature changes are estimated based on the local curvature tensor field. These local curvature changes are then transformed into curvature constraints. A sampling strategy is determined by combining the point cloud density, and shape quality indices that satisfy the local curvature changes are generated, including:

[0021] The local curvature tensor field contains curvature components of each point in the three coordinate axes. Based on the curvature components, eigenvalue equations are constructed. The principal curvature values ​​are obtained by solving the equations whose determinant is zero.

[0022] The Gaussian curvature is obtained by multiplying the principal curvature values, and the average curvature is obtained by calculating the arithmetic mean of the principal curvature values. The curvature gradient vector is obtained by calculating the partial derivatives of the Gaussian curvature and the average curvature in the three coordinate axis directions. The local curvature change rate is obtained by taking the square root of the sum of the squares of the components of the curvature gradient vector.

[0023] The local curvature change rate is compared with a preset curvature threshold, and the constraint coefficient is determined based on the comparison result. When the local curvature change rate meets the preset curvature threshold, the local curvature change rate is subtracted from the preset curvature threshold and then divided by a smoothing factor. The opposite number is used as the power of the exponential function to calculate the constraint coefficient.

[0024] Based on the constraint coefficients, a spherical search region is constructed in the point cloud data with each data point as the center. The number of points in the spherical search region is divided by the area of ​​the spherical search region to obtain the point density. The ratio of the standard deviation to the mean of the point density is calculated to obtain the shape quality index.

[0025] Based on the weld feature descriptor, welding current, voltage, and wire feed speed are used as joint optimization objectives. The optimal combination of process parameters and welding posture sequence are iteratively calculated using a multi-objective reward function. The mapping relationship between the optimal combination of process parameters and the welding posture sequence is established, including:

[0026] Based on the weld feature descriptor, the range of process parameter variation is determined, and process constraints between welding current, welding voltage, and wire feed speed are established. The welding current, welding voltage, and wire feed speed are then constructed as a joint optimization objective.

[0027] The penetration depth control factor is calculated based on the welding current, the welding voltage and the wire feeding speed, and the weld formation index is calculated based on the geometric parameters of the weld cross-section. The penetration depth control factor and the weld formation index are then used to construct a process quality evaluation index.

[0028] Collect welding torch posture angle and robot joint position data, calculate the deviation value of the welding torch posture angle relative to the target posture, calculate the motion smoothness of the robot joint position data, and construct the deviation value and the motion smoothness as motion performance evaluation index.

[0029] The process quality evaluation index and the motion performance evaluation index are combined into a multi-objective reward function. The joint optimization objective is iteratively optimized based on the gradient information of the multi-objective reward function. When the optimization gradient is less than a preset gradient threshold, the optimal combination of process parameters is output.

[0030] The welding torch posture sequence is optimized based on the optimal combination of process parameters, and a mapping relationship is established between the weld feature descriptor and the optimal combination of process parameters and the welding torch posture sequence, so as to realize the collaborative planning of process parameters and welding posture.

[0031] Calculate the deviation of the welding torch posture angle relative to the target posture, calculate the motion smoothness of the robot joint position, and construct a motion performance evaluation index from the deviation value and the motion smoothness, including:

[0032] Calculate the angular deviation between the welding torch attitude angle and the target attitude angle, multiply the angular deviation by the corresponding direction weight coefficient and sum them to obtain a weighted deviation index, which reflects the importance of different directional deviations;

[0033] The position and velocity signals of each joint of the robot are collected. The second derivative of the position signal is obtained to obtain the velocity difference signal. The absolute value of the velocity difference signal is calculated to obtain the position fluctuation index, which characterizes the acceleration change of the joint movement.

[0034] The velocity difference increment signal is obtained by taking the first derivative of the velocity signal, and the velocity change rate is obtained by calculating the absolute value of the velocity difference increment signal. The position fluctuation index and the velocity change rate are combined to obtain the motion stability index.

[0035] Based on the weighted deviation index and the motion stability index, a motion performance evaluation index is constructed using performance weighting coefficients, which are used to balance the influence of attitude accuracy and motion stability.

[0036] The mapping relationship is mapped to the robot joint space, and an improved fast collision detection algorithm is used for obstacle avoidance verification. The initial motion trajectory is generated by combining dynamic constraints, including:

[0037] The mapping relationships in the workspace are transformed to the robot joint space through inverse kinematics to obtain the joint angle sequence;

[0038] An improved fast collision detection algorithm is used for obstacle avoidance verification. The improved fast collision detection algorithm includes: constructing a hierarchical axial bounding box tree structure based on the joint angle sequence, updating the spatial position parameters of the axial bounding box tree structure in real time as the joint angle sequence changes; calculating the minimum distance between adjacent bounding boxes in the axial bounding box tree structure, comparing the minimum distance with a preset safety threshold, and marking a collision risk area when the minimum distance is less than the preset safety threshold.

[0039] A robot dynamics model is established, which includes a joint inertia matrix, a velocity coupling matrix, and a gravity term. Based on the robot dynamics model, position limits, velocity limits, acceleration limits, and driving torque limits of joint motion are set. The position limits, velocity limits, acceleration limits, and driving torque limits are used as constraints. An initial motion trajectory that satisfies the collision risk region is generated according to the constraints.

[0040] A piecewise adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided based on a comprehensive evaluation index of trajectory curvature, acceleration rate of change, and process requirements. Within the optimization interval, a variable structure filtering algorithm is used for trajectory smoothing. This dynamic optimization of the initial motion trajectory includes:

[0041] Calculate the curvature, acceleration, and process parameters of the initial motion trajectory; combine the curvature, acceleration, and process parameters into a comprehensive evaluation index; and construct an adaptive threshold function based on the comprehensive evaluation index.

[0042] The initial motion trajectory is divided into multiple optimization intervals based on the value of the adaptive threshold function. The adaptive threshold function decays exponentially with the change of the comprehensive evaluation index. The time point when the comprehensive evaluation index is greater than the value of the adaptive threshold function is determined as the interval boundary point.

[0043] A variable structure filter is established within the optimization interval based on the boundary points of the interval. The variable structure filter includes a state transition matrix and a state vector. A state estimate is calculated based on the observed values ​​of the state vector, and the state estimate is used as the optimized trajectory parameters.

[0044] Calculate the integral value of acceleration and the maximum deviation of trajectory parameters after optimization. When the integral value of acceleration and the maximum deviation of trajectory meet the preset parameter threshold, output a smooth trajectory. When they do not meet the preset parameter threshold, return to adjust the parameters of the variable structure filter until an optimized trajectory that meets the smoothness requirement is generated.

[0045] A second aspect of this invention provides a real-time optimization system for welding trajectories of industrial robots based on deep learning, comprising:

[0046] The first unit is used to collect three-dimensional point cloud data of complex curved surfaces, and to perform noise reduction and feature enhancement on the three-dimensional point cloud data using a multi-scale decomposition algorithm, and to generate high-precision curved surface data by combining a curvature-constrained mesh reconstruction method.

[0047] The second unit is used to establish a dynamic feature space on the high-precision surface data, extract the weld trajectory point set in the dynamic feature space, and construct a weld feature descriptor that integrates geometric features and topological relationships.

[0048] The third unit is used to take the welding current, voltage and wire feed speed as joint optimization objectives based on the weld feature descriptor, and to calculate the optimal combination of process parameters and welding posture sequence through multi-objective reward function iterative calculation, and to establish the mapping relationship between the optimal combination of process parameters and the welding posture sequence.

[0049] The fourth unit is used to map the mapping relationship to the robot joint space, perform obstacle avoidance verification using an improved fast collision detection algorithm, and generate the initial motion trajectory in combination with dynamic constraints.

[0050] The fifth unit is used to establish a segmented adaptive optimization model for the initial motion trajectory. It divides the optimization interval according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory.

[0051] The sixth unit is used to convert the optimized initial motion trajectory into robot control commands to achieve adaptive adjustment of the welding process.

[0052] A third aspect of the present invention provides an electronic device, comprising:

[0053] processor;

[0054] Memory used to store processor-executable instructions;

[0055] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0056] 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.

[0057] The beneficial effects of this application are as follows:

[0058] By collecting 3D point cloud data of complex curved surfaces and performing multi-scale decomposition and mesh reconstruction, high-precision curved surface data can be generated, which effectively improves the accuracy and stability of weld trajectory recognition and lays a solid foundation for subsequent welding trajectory planning.

[0059] By establishing a mapping relationship between process parameters and welding posture based on weld feature descriptors, and iteratively calculating the optimal parameter combination through a multi-objective reward function, adaptive optimization of process parameters can be achieved while ensuring welding quality, thereby improving welding efficiency and weld joint quality.

[0060] By employing a piecewise adaptive optimization model and a variable structure filtering algorithm to dynamically optimize the initial motion trajectory, and combining this with an improved fast collision detection algorithm for obstacle avoidance verification, a smooth and continuous welding trajectory that meets process requirements can be generated. This effectively reduces robot vibration and impact, extends equipment lifespan, and improves the stability and reliability of the welding process. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the real-time optimization method for industrial robot welding trajectory based on deep learning, as described in an embodiment of the present invention.

[0062] Figure 2 This is a flowchart of the point cloud curvature feature optimization based on multi-scale decomposition according to an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram illustrating the relationship between the number of point clouds and the reconstruction accuracy in an embodiment of the present invention;

[0064] Figure 4 This is a comparative analysis diagram of robot joint acceleration in an embodiment of the present invention;

[0065] Figure 5 This is a flowchart of the adaptive segmented filtering process for trajectory optimization in an embodiment of the present invention. Detailed Implementation

[0066] 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.

[0067] 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.

[0068] Figure 1This is a flowchart illustrating the real-time optimization method for industrial robot welding trajectories based on deep learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] Three-dimensional point cloud data of complex curved surfaces are collected, and the three-dimensional point cloud data are denoised and feature-enhanced using a multi-scale decomposition algorithm. High-precision curved surface data is generated by combining the curvature-constrained mesh reconstruction method.

[0070] A dynamic feature space is established on the high-precision surface data, and the weld trajectory point set in the dynamic feature space is extracted to construct a weld feature descriptor that integrates geometric features and topological relationships.

[0071] Based on the weld feature descriptor, welding current, voltage, and wire feed speed are used as joint optimization objectives. The optimal combination of process parameters and welding posture sequence are calculated iteratively through a multi-objective reward function, and the mapping relationship between the optimal combination of process parameters and the welding posture sequence is established.

[0072] The mapping relationship is mapped to the robot joint space, and an improved fast collision detection algorithm is used for obstacle avoidance verification. The initial motion trajectory is generated in combination with dynamic constraints.

[0073] A segmented adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory.

[0074] The optimized initial motion trajectory is converted into robot control commands to achieve adaptive adjustment of the welding process.

[0075] In one optional implementation, the three-dimensional point cloud data is denoised and feature-enhanced using a multi-scale decomposition algorithm, and high-precision surface data is generated by combining a curvature-constrained mesh reconstruction method, including:

[0076] The three-dimensional point cloud data is decomposed into multiple scales based on wavelet transform, and the local curvature change rate at each decomposition scale is calculated. The local curvature change rate is then used to perform multi-scale noise reduction on the three-dimensional point cloud data to obtain the noise-reduced point cloud data.

[0077] Calculate the local neighborhood covariance matrix for the denoised point cloud data, extract the eigenvalues ​​and eigenvectors of the local neighborhood covariance matrix, and construct a local geometric feature descriptor containing curvature information and normal vector information based on the eigenvalues ​​and eigenvectors.

[0078] The feature distance between the local geometric feature descriptors is calculated using the Gaussian kernel function. Based on the feature distance and the local point cloud density distribution, feature enhancement and density equalization are performed to generate feature-enhanced point cloud data.

[0079] A local curvature tensor field is constructed on the feature-enhanced point cloud data. The local curvature change is estimated based on the local curvature tensor field. The local curvature change is transformed into a curvature constraint condition. The sampling strategy is determined in combination with the point cloud density to generate a shape quality index that satisfies the local curvature change.

[0080] The shape quality index, the curvature constraint, and the point cloud fitting error are iteratively optimized and adjusted to adjust the mesh structure until convergence, at which point high-precision surface data is output.

[0081] The process involves acquiring 3D point cloud data, which may contain measurement noise and uneven sampling. Multi-scale decomposition of the 3D point cloud data is then performed based on wavelet transform. Specifically, 3D discrete wavelet transform is used to decompose the point cloud into different frequency components. For example, Haar wavelets or DB4 wavelets can be used as basis functions to decompose the point cloud coordinates into 3-5 levels, obtaining low-frequency approximate components and high-frequency detail components.

[0082] At each decomposition scale, the local curvature change rate is calculated, that is, within the neighborhood of point p (usually 10 to 20 nearest neighbor points are selected), the principal curvatures k1 and k2 are estimated by fitting the local quadratic surface, and then the curvature change rate is calculated as |Δk1|+|Δk2|, where Δk1 and Δk2 represent the curvature gradient along the principal direction, respectively.

[0083] Multi-scale denoising of point clouds is performed using the local rate of curvature change. Specifically, when the rate of curvature change of a point exceeds a threshold (e.g., twice the average value), the contribution of that point to the high-frequency components is attenuated by 50% to 80%. For regions with low and stable rates of curvature change, their high-frequency detail information is preserved. The processed frequency components are then reconstructed using inverse wavelet transform to obtain the denoised point cloud data.

[0084] Calculate the local neighborhood covariance matrix for the denoised point cloud data. For each point p, select neighborhood points within a radius r (r is usually 2 to 3 times the average sampling distance of the point cloud) to construct a 3×3 covariance matrix. Extract the eigenvalues ​​λ1≥λ2≥λ3 and the corresponding eigenvectors v1, v2, and v3 of this covariance matrix, where v3 approximately represents the normal vector at point p.

[0085] A local geometric feature descriptor is constructed based on eigenvalues ​​and eigenvectors. This descriptor includes: surface variation index λ3 / (λ1+λ2+λ3), principal curvature ratio (λ1-λ2) / (λ1+λ2), normal vector v3, and curvature variation along the principal direction. These features together constitute an 8- to 12-dimensional feature vector, used to describe the local geometric properties at point p.

[0086] The Gaussian kernel function is used to calculate the feature distance between local geometric feature descriptors. For points p and q, the feature distance d(p,q) = exp(-||fp-fq||² / σ²), where fp and fq are the feature descriptors of the two points, and σ is the kernel width parameter (usually taken as 0.6 to 0.8 times the average distance of the feature vectors).

[0087] Feature enhancement and density equalization are performed based on the feature distance and local point cloud density distribution: In regions with small feature distances (e.g., less than 0.2), new points are generated through interpolation to enhance model details; in regions with low point cloud density but significant geometric features (e.g., edges, corners), the sampling point density is increased to 1.5–2 times the average density; in flat regions, the point density is appropriately reduced to 0.7–0.8 times the average density. This process generates feature-enhanced point cloud data that maintains the original geometric features while having a more balanced point distribution.

[0088] A local curvature tensor field is constructed on feature-enhanced point cloud data. For each point p, a quadratic surface is fitted based on its neighborhood points, and the principal curvatures k1 and k2 and their corresponding principal directions are calculated. Using tensor interpolation, the discrete curvature information is extended into a continuous curvature tensor field. Local curvature changes are estimated based on the local curvature tensor field, and the curvature gradient magnitude and direction are calculated. The curvature changes are then transformed into curvature constraints: in high curvature regions (e.g., k1 or k2 is greater than 3 times the average curvature), the grid edge length should not exceed 0.5 times the average sampling distance; in regions with drastic curvature changes (curvature gradient magnitude is greater than a threshold), the grid surface should be subdivided along the direction of maximum curvature change.

[0089] The sampling strategy is determined by combining point cloud density. While ensuring mesh quality, sampling can be appropriately reduced in high-density areas, and sampling points can be increased in low-density areas through interpolation. This generates shape quality indicators that satisfy local curvature changes, including: mesh area uniformity, mesh side length and local curvature fit, and triangle interior angle size (ideally close to 60°).

[0090] The mesh structure is adjusted iteratively by considering shape quality indices, curvature constraints, and point cloud fitting errors. The specific optimization process employs an energy minimization method: the energy function is defined as E = w1Eshape + w2Ecurvature + w3Efit, where Eshape represents the shape quality index of the mesh structure, Ecurvature represents the curvature constraint, Efit represents the point cloud fitting error, and w1, w2, and w3 are weighting coefficients (typically 0.4, 0.3, and 0.3). In each iteration, operations such as edge flipping, vertex movement, and face splitting are used to adjust the mesh and reduce the energy function value.

[0091] When the rate of change of the energy function is less than a preset threshold (e.g., 0.001) or the maximum number of iterations (e.g., 50) is reached, the optimization process converges, outputting high-precision surface data. This surface data is represented in the form of a triangular mesh, while preserving the geometric features and curvature distribution information of the point cloud.

[0092] Experiments show that for point cloud data of mechanical parts containing 5% Gaussian noise (approximately 500,000 points), after processing with this method, the root mean square error of the surface is reduced from the original 0.35 mm to 0.08 mm, the curvature fidelity is improved by 78%, and the geometric features in the edge and detail areas are effectively preserved.

[0093] Figure 2 Here is a flowchart of the point cloud curvature feature optimization based on multi-scale decomposition according to an embodiment of the present invention:

[0094] This flowchart details a complete technical solution for point cloud data processing and optimization, comprising five key steps. First, wavelet transform is used to decompose the 3D point cloud data into multiple scales. Noise reduction is achieved by calculating the rate of change of local curvature at different decomposition scales. Second, the local neighborhood covariance matrix is ​​calculated for the denoised point cloud data, extracting eigenvalues ​​and eigenvectors. Based on these features, local geometric feature descriptors containing curvature and normal vector information are constructed. Third, the feature distance between feature descriptors is calculated using a Gaussian kernel function, and feature enhancement and density equalization are performed in conjunction with the local point cloud density distribution. Fourth, a local curvature tensor field is constructed based on the feature-enhanced point cloud data. Local curvature changes are calculated and transformed into curvature constraints, and a sampling strategy is determined based on the point cloud density. Finally, the shape quality index, curvature constraints, and point cloud fitting error are iteratively optimized and adjusted to adjust the mesh structure until convergence is achieved, outputting high-precision surface data. This multi-step optimization method effectively improves the quality and accuracy of point cloud data.

[0095] In one optional implementation, a local curvature tensor field is constructed on the feature-enhanced point cloud data. Local curvature changes are estimated based on the local curvature tensor field. These local curvature changes are then transformed into curvature constraints. A sampling strategy is determined by combining the point cloud density, and a shape quality index satisfying the local curvature changes is generated, including:

[0096] The local curvature tensor field contains curvature components of each point in the three coordinate axes. Based on the curvature components, eigenvalue equations are constructed. The principal curvature values ​​are obtained by solving the equations whose determinant is zero.

[0097] The Gaussian curvature is obtained by multiplying the principal curvature values, and the average curvature is obtained by calculating the arithmetic mean of the principal curvature values. The curvature gradient vector is obtained by calculating the partial derivatives of the Gaussian curvature and the average curvature in the three coordinate axis directions. The local curvature change rate is obtained by taking the square root of the sum of the squares of the components of the curvature gradient vector.

[0098] The local curvature change rate is compared with a preset curvature threshold, and the constraint coefficient is determined based on the comparison result. When the local curvature change rate meets the preset curvature threshold, the local curvature change rate is subtracted from the preset curvature threshold and then divided by a smoothing factor. The opposite number is used as the power of the exponential function to calculate the constraint coefficient.

[0099] Based on the constraint coefficients, a spherical search region is constructed in the point cloud data with each data point as the center. The number of points in the spherical search region is divided by the area of ​​the spherical search region to obtain the point density. The ratio of the standard deviation to the mean of the point density is calculated to obtain the shape quality index.

[0100] The system receives feature-enhanced point cloud data. This feature-enhanced point cloud data contains the three-dimensional coordinate information of each point and the surface normal vector information of each point. The system constructs a local curvature tensor field on this point cloud data, which describes the curvature of the point cloud surface in different directions.

[0101] For each point in the point cloud, the system constructs a local curvature tensor by analyzing the distribution of its neighboring points. Specifically, with the current point as the center, all points within a spherical region with a radius of 0.05 meters are selected as neighboring points. For these neighboring points, the system calculates their relative position vectors with respect to the center point and then constructs a covariance matrix.

[0102] By performing eigenvalue decomposition on the covariance matrix, three eigenvectors and their corresponding eigenvalues ​​are obtained. These three eigenvectors represent the principal directions of the local surface, while the eigenvalues ​​reflect the curvature components in these directions. For example, for a point in a planar region, one of the eigenvalues ​​is close to zero; while for a point in a spherical region, the three eigenvalues ​​are of similar magnitude.

[0103] Based on the curvature components calculated above, the system constructs an eigenvalue equation. This equation is represented as a third-order determinant equal to zero. By solving this equation, the system obtains three eigenvalues, of which the two largest eigenvalues ​​are defined as the principal curvature values. For example, for a certain point, solving the eigenvalue equation may yield eigenvalues ​​0.1, 0.08, and 0.01; then 0.1 and 0.08 are the principal curvature values ​​at that point.

[0104] After obtaining the principal curvature values, the system calculates the Gaussian curvature and the mean curvature. The Gaussian curvature is the product of the two principal curvature values, and the mean curvature is the arithmetic mean of the two principal curvature values. Taking the example above, the Gaussian curvature at this point is 0.1 multiplied by 0.08, which equals 0.008, and the mean curvature is (0.1 plus 0.08) divided by 2, which equals 0.09.

[0105] To assess the changes in curvature, the system calculates the partial derivatives of the Gaussian curvature and the mean curvature along the three coordinate axes, forming curvature gradient vectors. Specifically, the system selects neighboring points within 0.01 meters of the current point in the x, y, and z directions, calculates the Gaussian curvature and mean curvature of these points, and then divides the difference between these values ​​and the curvature value of the current point by the distance to obtain the curvature gradient in each direction. For example, the Gaussian curvature gradient of a point might be 0.002 in the x-direction, 0.003 in the y-direction, and 0.001 in the z-direction.

[0106] The local rate of change of curvature is obtained by taking the square root of the sum of the squares of the components of the curvature gradient vector. Continuing with the example above, the local rate of change of curvature at this point is the square root of (0.002 squared plus 0.003 squared plus 0.001 squared), which is approximately equal to 0.0037.

[0107] The calculated rate of change of local curvature is compared with a preset curvature threshold, and the constraint coefficient is determined based on the comparison result. The preset curvature threshold can be set to 0.003. When the rate of change of local curvature is greater than the preset curvature threshold, the system subtracts the preset curvature threshold from the rate of change of local curvature, then divides by a smoothing factor (such as 0.001), takes the negative of the result as the power of the exponential function, and calculates the constraint coefficient.

[0108] Based on the calculated constraint coefficients, the system constructs a spherical search region centered on each data point in the point cloud data. The smaller the constraint coefficient, the larger the radius of the spherical search region. For example, for a constraint coefficient of 0.497, the system might set the radius of the spherical search region to 0.1 meters; while for points with constraint coefficients close to 1, the system might set the radius to 0.05 meters.

[0109] Count the number of points within the spherical search area and divide the number of points by the surface area of ​​the spherical search area to obtain the point density. For example, if there are 314 points in a spherical area with a radius of 0.1 meters, then the point density is 314 divided by (4π multiplied by 0.1 squared), which is approximately 250 points per square meter.

[0110] The shape quality index is calculated by dividing the standard deviation of the point density by the mean. For example, if the mean point density is 200 points per square meter and the standard deviation is 40 points per square meter, then the shape quality index is 40 divided by 200, which equals 0.2. This index reflects the uniformity of point cloud sampling and its adaptability to local curvature variations. A lower shape quality index indicates more uniform point cloud sampling and better adaptation to curvature variations.

[0111] Through the above steps, the system can adaptively adjust the sampling strategy according to the local curvature characteristics of the point cloud, increasing the sampling density in areas with large curvature changes and decreasing the sampling density in areas with small curvature changes, thereby generating shape quality indicators that meet the local curvature changes, providing a basis for subsequent point cloud processing and analysis.

[0112] Figure 3 This is a schematic diagram illustrating the relationship between the number of point clouds and the reconstruction accuracy in an embodiment of the present invention:

[0113] This figure compares the accuracy performance of three different methods in point cloud reconstruction. The horizontal axis represents the number of points (from 102 to 107), and the vertical axis represents the reconstruction accuracy error (in mm). The three methods are represented by lines of different colors: our proposed method (triangles), Poisson reconstruction (dots), and RANSAC fitting (squares). The data trend shows that the reconstruction error of all three methods decreases as the number of points increases. Our proposed method performs best, with the error decreasing from 0.032 mm with 102 points to 0.006 mm with 107 points. The Poisson reconstruction method has the largest error, decreasing from 0.068 mm with 102 points to 0.029 mm with 107 points. The RANSAC fitting method falls in between, decreasing from 0.053 mm with 102 points to 0.019 mm with 107 points. Overall, this technical solution demonstrates better stability and lower error levels as the number of point clouds increases. In particular, when the number of point clouds reaches more than 105, its advantage in reconstruction accuracy becomes more obvious, indicating that the solution has better adaptability and accuracy when processing large-scale point cloud data.

[0114] In one optional implementation, based on the weld feature descriptor, welding current, voltage, and wire feed speed are used as joint optimization objectives. The optimal combination of process parameters and welding posture sequence are iteratively calculated using a multi-objective reward function. Establishing the mapping relationship between the optimal combination of process parameters and the welding posture sequence includes:

[0115] Based on the weld feature descriptor, the range of process parameter variation is determined, and process constraints between welding current, welding voltage, and wire feed speed are established. The welding current, welding voltage, and wire feed speed are then constructed as a joint optimization objective.

[0116] The penetration depth control factor is calculated based on the welding current, the welding voltage and the wire feeding speed, and the weld formation index is calculated based on the geometric parameters of the weld cross-section. The penetration depth control factor and the weld formation index are then used to construct a process quality evaluation index.

[0117] Collect welding torch posture angle and robot joint position data, calculate the deviation value of the welding torch posture angle relative to the target posture, calculate the motion smoothness of the robot joint position data, and construct the deviation value and the motion smoothness as motion performance evaluation index.

[0118] The process quality evaluation index and the motion performance evaluation index are combined into a multi-objective reward function. The joint optimization objective is iteratively optimized based on the gradient information of the multi-objective reward function. When the optimization gradient is less than a preset gradient threshold, the optimal combination of process parameters is output.

[0119] The welding torch posture sequence is optimized based on the optimal combination of process parameters, and a mapping relationship is established between the weld feature descriptor and the optimal combination of process parameters and the welding torch posture sequence, so as to realize the collaborative planning of process parameters and welding posture.

[0120] Based on the weld feature descriptor, the range of process parameter variation is determined, and process constraints between welding current, welding voltage, and wire feed speed are established. Welding current, welding voltage, and wire feed speed are then constructed as a joint optimization objective.

[0121] For V-type welds, the characteristic descriptor includes bevel angle, blunt edge dimension, and plate thickness information. When the plate thickness is 10mm, the bevel angle is 60°, and the blunt edge dimension is 2mm, based on empirical data for Q235 steel, the welding current range is determined to be 180A to 240A, the welding voltage range to be 22V to 28V, and the wire feed speed range to be 6m / min to 10m / min. Process constraints are established: the welding current is directly proportional to the wire feed speed, i.e., for every 1m / min increase in wire feed speed, the welding current increases by approximately 15A; the welding voltage is weakly correlated with the welding current, with every 1V increase in voltage increasing the welding current by approximately 5A. These process constraints form a three-dimensional parameter space, serving as the search range for joint optimization.

[0122] The penetration depth control factor is calculated based on welding current, welding voltage, and wire feed speed. The weld formation index is calculated based on the geometric parameters of the weld cross-section. These two factors are then used to construct a process quality evaluation index. The penetration depth control factor is calculated by dividing the product of welding current and welding voltage by the wire feed speed.

[0123] For example, when the welding current is 220A, the welding voltage is 25V, and the wire feed speed is 8m / min, the penetration control factor is 687.5. The larger this value, the deeper the penetration. Practice shows that for a 10mm thick V-type weld, the ideal range for the penetration control factor is 650 to 700. The weld formation index is measured by the ratio of weld reinforcement to weld width; the ideal weld formation index should be controlled between 0.1 and 0.2. When the welding parameters are within the above range, and the actual measured weld reinforcement is 2mm and the weld width is 15mm, the weld formation index is 0.133, which is within the ideal range.

[0124] Collect welding torch posture angle and robot joint position data, calculate the deviation value of welding torch posture angle relative to the target posture, calculate the motion smoothness of robot joint position data, and construct motion performance evaluation index by combining deviation value and motion smoothness.

[0125] The welding torch attitude angles are obtained by real-time acquisition of the rotation matrix at the end of the welding torch, including the feed angle and the oscillation angle. For V-shaped welds, the ideal feed angle is 15° and the oscillation angle is 0°. When the actual feed angle is 18° and the oscillation angle is 2°, the attitude deviation is 5° (calculated using Euclidean distance). Robot joint position data includes the angle values ​​of six joints. The smoothness of joint movement is obtained by calculating the rate of change of joint angular velocity at adjacent time points. Experimental data shows that when the rate of change of joint angular velocity is less than 5° / s², the welding motion smoothness is good, corresponding to an evaluation value of 0.9 (out of 1).

[0126] A multi-objective reward function is constructed by combining process quality evaluation indicators and motion performance evaluation indicators. The joint optimization objective is iteratively optimized based on the gradient information of the multi-objective reward function. The optimal combination of process parameters is output when the optimization gradient is less than a preset gradient threshold. The multi-objective reward function is constructed as follows: the evaluation value of the weld penetration control factor (maximum score 1) is multiplied by a weight of 0.4; the evaluation value of the weld formation index (maximum score 1) is multiplied by a weight of 0.3; the evaluation value of the attitude deviation value (maximum score 1) is multiplied by a weight of 0.2; and the evaluation value of motion stability is multiplied by a weight of 0.1. These four values ​​are then weighted and summed to form the final reward value. For example, R = 0.4P. d +0.3P f +0.2P p +0.1P s Where R is the final comprehensive reward value, ranging from [0,1]; P d P is the evaluation value of the melt depth control factor; f P is the evaluation value of the weld formation index. p The evaluation value representing the attitude deviation; P s The evaluation value represents the smoothness of motion.

[0127] The formula for calculating the evaluation value of the melt depth control factor is as follows:

[0128] P d =exp(-|h a -h d | / h d );

[0129] Among them, h a The actual penetration depth is calculated from the current (I), voltage (U), and wire feed speed (v), h d The target melting depth value;

[0130] The formula for calculating the weld formation index is as follows:

[0131] P f =exp(-|w / h-α| / α);

[0132] Where w is the weld width, h is the weld height, and α is the ideal weld width-to-height ratio (usually 1.2-1.5);

[0133] The formula for calculating the evaluation value of attitude deviation is as follows:

[0134] P p =exp(-θ / θ m ax);

[0135] Where θ is the angle between the welding torch and the workpiece normal, θ m ax is the maximum allowable deviation angle (usually 15°);

[0136] The formula for calculating the evaluation value of motion smoothness is as follows:

[0137] P s =exp(-|a| / a m ax);

[0138] Where a is the acceleration of the welding torch, a m ax is the maximum allowable acceleration value.

[0139] The optimization process employed gradient descent for iterative calculations, with a preset gradient threshold of 0.01. After approximately 50 iterations, the optimal combination of process parameters was obtained: welding current of 225A, welding voltage of 24.5V, and wire feed speed of 8.5m / min. At this point, the penetration control factor was 650.7, the weld formation index was 0.15, the attitude deviation was 2°, the motion stability was 0.95, and the total bonus was 0.92.

[0140] The welding torch posture sequence is optimized based on the optimal combination of process parameters. A mapping relationship is established from the weld feature descriptor to the optimal combination of process parameters and the welding torch posture sequence, realizing the collaborative planning of process parameters and welding posture. The method for optimizing the welding torch posture sequence is as follows: with the optimal process parameters fixed, the posture combination that makes the weld pool most stable is found by adjusting three parameters: the feed angle, the oscillation angle, and the welding torch height. For the above V-shaped weld, the optimal welding torch posture parameters are: feed angle 15°, oscillation angle 0°, and welding torch height 12mm.

[0141] The attitude parameters are evenly distributed along the weld path to form a welding attitude sequence. The final established mapping relationship is as follows: when the weld feature descriptor is {plate thickness = 10mm, bevel angle = 60°, blunt edge dimension = 2mm}, the corresponding optimal process parameter combination is {welding current = 225A, welding voltage = 24.5V, wire feed speed = 8.5m / min}, and the corresponding optimal welding attitude sequence is {feed angle = 15°, oscillation angle = 0°, welding torch height = 12mm}. This mapping relationship can be stored in a database to guide the welding of similar feature welds.

[0142] In one optional implementation, the deviation value of the welding torch posture angle relative to the target posture is calculated, the motion smoothness of the robot joint position is calculated, and the deviation value and the motion smoothness are used to construct a motion performance evaluation index, including:

[0143] Calculate the angular deviation between the welding torch attitude angle and the target attitude angle, multiply the angular deviation by the corresponding direction weight coefficient and sum them to obtain a weighted deviation index, which reflects the importance of different directional deviations;

[0144] The position and velocity signals of each joint of the robot are collected. The second derivative of the position signal is obtained to obtain the velocity difference signal. The absolute value of the velocity difference signal is calculated to obtain the position fluctuation index, which characterizes the acceleration change of the joint movement.

[0145] The velocity difference increment signal is obtained by taking the first derivative of the velocity signal, and the velocity change rate is obtained by calculating the absolute value of the velocity difference increment signal. The position fluctuation index and the velocity change rate are combined to obtain the motion stability index.

[0146] Based on the weighted deviation index and the motion stability index, a motion performance evaluation index is constructed using performance weighting coefficients, which are used to balance the influence of attitude accuracy and motion stability.

[0147] The welding robot's end-effector position sensor collects real-time information on the welding torch's actual attitude angles, including pitch, roll, and yaw. Simultaneously, the target attitude angle at the corresponding moment is obtained from the welding trajectory planning module. For each attitude angle component, the angle difference between the actual and target values ​​is calculated, yielding the angle deviations in three directions. For example, assuming at a certain moment, the actual pitch angle of the welding torch is 32.5 degrees and the target pitch angle is 30.0 degrees, the pitch angle deviation is 2.5 degrees; the actual roll angle is 45.2 degrees and the target roll angle is 45.0 degrees, the roll angle deviation is 0.2 degrees; and the actual yaw angle is 88.7 degrees and the target yaw angle is 90.0 degrees, the yaw angle deviation is 1.3 degrees.

[0148] Because attitude deviations in different directions have varying impacts on welding quality during the welding process, directional weighting coefficients are needed. Based on the weld type and process requirements, the pitch angle weighting coefficient is set to 0.5, the roll angle weighting coefficient to 0.3, and the yaw angle weighting coefficient to 0.2. The weighted deviation index is obtained by multiplying each directional angle deviation by its corresponding weighting coefficient and summing the results. Using the above data as an example, the weighted deviation index is calculated as: 2.5 × 0.5 + 0.2 × 0.3 + 1.3 × 0.2 = 1.51. This index comprehensively reflects the overall degree of deviation in the welding torch attitude, while also highlighting the differences in the importance of deviations in different directions.

[0149] The motion smoothness evaluation of robot joint positions consists of two parts: position fluctuation index and velocity change rate. For the position fluctuation index, the system collects position signals from the robot's six joints at a frequency of 100Hz, denoted as the position sequences of joints 1 to 6. The second derivative is calculated for each joint position sequence; that is, the velocity sequence is obtained by first calculating the position difference between two adjacent points, and then the velocity difference signal is obtained by calculating the difference between the velocity sequences. The absolute value of the velocity difference signal is taken, and the average value is calculated as the position fluctuation index for that joint. For example, the position fluctuation indices for the six robot joints are 0.024, 0.015, 0.032, 0.008, 0.012, and 0.005 (unit: degrees / second²).

[0150] For the velocity change rate index, the system directly collects the velocity signals of each joint of the robot and calculates the first derivative of the velocity signals to obtain the velocity difference increment signal. After taking the absolute value of the velocity difference increment signal, the average value is calculated as the velocity change rate. For example, the velocity change rates of the robot's six joints are 0.118, 0.092, 0.145, 0.056, 0.074, and 0.038 (unit: degrees / second²).

[0151] The motion smoothness index is obtained by combining the position fluctuation index and the velocity change rate. Specifically, the position fluctuation index of the six joints is averaged to obtain 0.016; the velocity change rate of the six joints is averaged to obtain 0.087. These two indices are then combined in a 7:3 ratio to calculate the motion smoothness index as 0.016 × 0.7 + 0.087 × 0.3 = 0.0372. The smaller this index value, the smoother the robot's motion.

[0152] The final motion performance evaluation index is constructed by combining a weighted deviation index and a motion stability index. With a weighting coefficient of 0.6 for attitude accuracy and 0.4 for motion stability, the motion performance evaluation index is calculated as: 1.51 × 0.6 + 0.0372 × 0.4 = 0.921. This comprehensive index provides a complete evaluation of the welding robot's motion performance, and the weighting coefficients reflect a balanced consideration of attitude accuracy and motion stability.

[0153] The weighting coefficients can be adjusted for different types of welding tasks. For example, for precision welding tasks, the weight of attitude accuracy in the final evaluation index can be increased to 0.8; for continuous long weld welding, the weight of motion stability can be increased to 0.6. The system also provides a dynamic weighting adjustment function, which automatically adjusts the weighting coefficients based on real-time feedback during the welding process to further optimize the evaluation results.

[0154] In practical applications, the system optimizes the welding trajectory based on this evaluation index. When the evaluation index exceeds the preset threshold of 1.2, the trajectory optimization module is triggered, which improves the trajectory quality by adjusting path planning parameters or adding intermediate points. The optimized trajectory reduces the evaluation index from 1.35 to 0.87, significantly improving welding quality and reducing the weld defect rate by 32%.

[0155] Furthermore, this evaluation method supports both offline analysis and online monitoring modes. Offline analysis is used for trajectory evaluation during the welding program development stage, while online monitoring is used for real-time monitoring during the production process, providing a basis for dynamic adjustment of welding parameters and achieving comprehensive quality assurance of the welding process.

[0156] Figure 4 A comparative analysis diagram of robot joint acceleration in an embodiment of the present invention:

[0157] This figure compares the acceleration performance of three different schemes at six joints of the robot. The horizontal axis represents the joint numbers of the robot (from joint 1 to joint 6), and the vertical axis represents the rate of change of acceleration (rad / s). 3 The three schemes are represented by different lines: this technical scheme (triangle), traditional trajectory planning (dot), and fifth-order polynomial interpolation (square). From the data trend, this technical scheme exhibits the lowest acceleration value at all joint positions, starting from 0.032 rad / s² at joint 1. 3 Gradually decrease to 0.015 rad / s at joint 6. 3 The traditional trajectory planning scheme produces the highest acceleration value, starting from 0.085 rad / s² at joint 1. 3 Reduced to 0.060 rad / s at joint 6. 3 The performance of the quintic polynomial interpolation scheme falls between the two, from 0.052 rad / s at joint 1. 3 Reduced to 0.035 rad / s at joint 6. 3 Overall, the acceleration values ​​of the three schemes all show a trend of gradually decreasing from joint 1 to joint 6. However, the technical scheme of this invention shows a significant advantage in reducing joint acceleration, which helps to improve the smoothness and accuracy of robot motion.

[0158] In one optional implementation, the mapping relationship is mapped to the robot joint space, an improved fast collision detection algorithm is used for obstacle avoidance verification, and the initial motion trajectory is generated in combination with dynamic constraints, including:

[0159] The mapping relationships in the workspace are transformed to the robot joint space through inverse kinematics to obtain the joint angle sequence;

[0160] An improved fast collision detection algorithm is used for obstacle avoidance verification. The improved fast collision detection algorithm includes: constructing a hierarchical axial bounding box tree structure based on the joint angle sequence, updating the spatial position parameters of the axial bounding box tree structure in real time as the joint angle sequence changes; calculating the minimum distance between adjacent bounding boxes in the axial bounding box tree structure, comparing the minimum distance with a preset safety threshold, and marking a collision risk area when the minimum distance is less than the preset safety threshold.

[0161] A robot dynamics model is established, which includes a joint inertia matrix, a velocity coupling matrix, and a gravity term. Based on the robot dynamics model, position limits, velocity limits, acceleration limits, and driving torque limits of joint motion are set. The position limits, velocity limits, acceleration limits, and driving torque limits are used as constraints. An initial motion trajectory that satisfies the collision risk region is generated according to the constraints.

[0162] In a robot motion planning method, it is necessary to transform the mapping relationships in the workspace to the joint space, perform collision detection and obstacle avoidance verification, and generate the initial motion trajectory while considering dynamic constraints. Specifically, this method employs the following techniques:

[0163] For the mapping transformation between the workspace and joint space, the system converts the sequence of path points in the workspace into a sequence of joint angles using inverse kinematics. In practical applications, assuming the end effector of a six-axis robotic arm needs to move from position point P1 (500mm, 300mm, 200mm) to position point P2 (600mm, 350mm, 250mm), the system first generates an intermediate path point between these two points, such as P_mid (550mm, 325mm, 225mm), forming the path point sequence {P1, P_mid, P2}. Then, the inverse kinematics algorithm is applied to each path point to obtain the corresponding sequence of joint angles. For example, the joint angles corresponding to point P1 are {30°, 45°, -20°, 15°, 60°, 10°}, the joint angles corresponding to point P_mid are {35°, 48°, -18°, 18°, 62°, 12°}, and the joint angles corresponding to point P2 are {40°, 50°, -15°, 20°, 65°, 15°}, thus forming a complete joint angle sequence.

[0164] For collision detection and obstacle avoidance verification, this method employs an improved fast collision detection algorithm. This algorithm first constructs a hierarchical axial bounding box (AABB) tree structure based on joint angle sequences. In practical implementation, the system represents each link of the robot as a basic AABB, and then organizes these basic AABBs into a tree structure. For example, for a six-axis robotic arm, the base corresponds to AABB_0, the first link corresponds to AABB_1, and so on until the end effector corresponds to AABB_6. Each AABB is represented by the coordinates of its minimum and maximum vertices; for example, the parameters of AABB_1 might be {minimum point (-100mm, -50mm, 0mm), maximum point (100mm, 50mm, 300mm)}.

[0165] As the joint angle sequence changes, the system updates the spatial position parameters of the AABB tree structure in real time. When the joint angle changes from {30°, 45°, -20°, 15°, 60°, 10°} to {35°, 48°, -18°, 18°, 62°, 12°}, the system calculates the new position and orientation of each link based on forward kinematics and updates the spatial position parameters of the corresponding AABB accordingly. For example, AABB_2 may be updated to {minimum point (-80mm, -60mm, 280mm), maximum point (120mm, 40mm, 580mm)} under the new joint angle.

[0166] During collision detection, the system calculates the minimum distance between adjacent bounding boxes in the AABB tree structure. Specifically, for any two AABBs, the system first determines if they overlap; if not, it calculates their minimum distance. For example, the minimum distance between AABB_2 and environmental obstacles AABB_obs{minimum point (200mm, -100mm, 300mm), maximum point (300mm, 100mm, 500mm)} is calculated to be 80mm. The system compares this minimum distance with a preset safety threshold of 100mm. Since 80mm is less than 100mm, this location is marked as a collision risk area.

[0167] To generate an initial motion trajectory that satisfies the dynamic constraints, the system first establishes a robot dynamics model that includes joint inertia matrices, velocity coupling matrices, and gravity terms. In practical applications, the system can obtain the mass, center of mass position, and inertia tensor of each link through the robot's CAD parameters. For example, the mass of the first link is 5 kg, the center of mass position is (0 mm, 0 mm, 150 mm), and the principal axis moment of inertia is (0.3 kg·m², 0.3 kg·m², 0.1 kg·m²).

[0168] Based on the dynamic model, the system sets various limit constraints for joint motion. For position limits, for example, the range of motion of joint 1 is set to [-170°, 170°]; the speed limit is set to the maximum rotational speed of each joint, such as 120° / s for joint 1; the acceleration limit is set to the maximum acceleration of each joint, such as 180° / s² for joint 1; and the driving torque limit is set to the maximum output torque of each motor, such as 150 N·m for joint 1.

[0169] After marking the collision risk areas, the system generates an initial trajectory based on the aforementioned constraints. For locations marked as collision risk areas, the system modifies the original path points and adds obstacle avoidance points.

[0170] For example, in the original path P1→P_mid→P2, if there is a collision risk at P_mid, the system will generate a new obstacle avoidance point P_avoid(550mm,325mm,300mm), making the new path P1→P_avoid→P2.

[0171] The new path is converted into a sequence of joint angles, and its compliance with all dynamic constraints is verified. If some points do not meet the constraints, the system adjusts the time parameters, reducing the velocity and acceleration until all constraints are met, ultimately generating a feasible initial motion trajectory.

[0172] By comprehensively applying the above technical means, this method can effectively transform the workspace mapping relationship into the joint space, realize rapid collision detection and obstacle avoidance, and generate a safe and feasible initial motion trajectory for the robot while satisfying dynamic constraints.

[0173] In one optional implementation, a piecewise adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided based on a comprehensive evaluation index of trajectory curvature, acceleration rate of change, and process requirements. Within the optimization interval, a variable structure filtering algorithm is used for trajectory smoothing, thereby achieving dynamic optimization of the initial motion trajectory.

[0174] Calculate the curvature, acceleration, and process parameters of the initial motion trajectory; combine the curvature, acceleration, and process parameters into a comprehensive evaluation index; and construct an adaptive threshold function based on the comprehensive evaluation index.

[0175] The initial motion trajectory is divided into multiple optimization intervals based on the value of the adaptive threshold function. The adaptive threshold function decays exponentially with the change of the comprehensive evaluation index. The time point when the comprehensive evaluation index is greater than the value of the adaptive threshold function is determined as the interval boundary point.

[0176] A variable structure filter is established within the optimization interval based on the boundary points of the interval. The variable structure filter includes a state transition matrix and a state vector. A state estimate is calculated based on the observed values ​​of the state vector, and the state estimate is used as the optimized trajectory parameters.

[0177] Calculate the integral value of acceleration and the maximum deviation of trajectory parameters after optimization. When the integral value of acceleration and the maximum deviation of trajectory meet the preset parameter threshold, output a smooth trajectory. When they do not meet the preset parameter threshold, return to adjust the parameters of the variable structure filter until an optimized trajectory that meets the smoothness requirement is generated.

[0178] A segmented adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory.

[0179] Acquire initial motion trajectory data, including position coordinates, velocity, acceleration, and other parameters at each time point. Calculate the curvature value of the initial motion trajectory by determining the radius of curvature using a circle formed by three adjacent points, and taking its reciprocal as the curvature value. The rate of change of acceleration is obtained by calculating the derivative of acceleration. Process parameters are determined according to the application scenario, such as feed rate and spindle speed in machining. Combine the curvature K, the rate of change of acceleration A, and the process parameter P to form a comprehensive evaluation index Q, calculated as follows:

[0180] Q = w1×K + w2×A + w3×P, where w1, w2, and w3 are weighting coefficients that can be adjusted according to the actual application. For example, in a precision machining scenario, w1 = 0.4, w2 = 0.5, and w3 = 0.1 can be set.

[0181] An adaptive threshold function T is constructed based on a comprehensive evaluation index, T = T0 × e^(-λ × Q), where T0 is the initial threshold and λ is the attenuation coefficient. Taking the motion trajectory of a precision machining center as an example, the initial threshold T0 is set to 5.0, and the attenuation coefficient λ is set to 0.2. When the comprehensive evaluation index Q in the region of drastic curvature change reaches 3.5, the corresponding threshold function value T drops to 1.84, while in the straight line region when Q is 0.5, the threshold T remains at 4.09, realizing the adaptive change of the threshold with the motion state.

[0182] The initial trajectory is divided into multiple optimization intervals based on a comprehensive evaluation index. The comprehensive evaluation index Q at each time point is compared with the corresponding threshold function value T. When Q > T, the time point is marked as the interval boundary point. For example, the times 0s, 2.5s, 4.8s, 7.2s, and 10s of the initial trajectory are identified as boundary points, thus dividing the 10-second trajectory into 4 optimization intervals.

[0183] A variable structure filter is established within each optimization interval. The filter state transition matrix is ​​constructed based on kinematic relationships. For a two-dimensional planar trajectory, the state vector includes position coordinates x and y, their first derivatives vx and vy, and their second derivatives ax and ay. The filter parameters are adaptively adjusted according to the interval characteristics. Intervals with greater curvature have higher position weights, while intervals with greater velocity changes have higher velocity weights. Taking the 0-2.5s interval as an example, this interval has greater curvature, so the position weight is set to 0.6, the velocity weight to 0.3, and the acceleration weight to 0.1; while in the straight-line interval of 4.8-7.2s, the position weight is set to 0.4, the velocity weight to 0.5, and the acceleration weight to 0.1.

[0184] Using the sampled points of the initial trajectory as observations, the state estimate is iteratively calculated through a variable structure filter. For each time point t, the current state is predicted based on the state estimate from the previous time point t-1 and the state transition matrix. This prediction is then combined with the observations to update the state estimate for greater accuracy. The state estimate includes position, velocity, and acceleration, which serve as the optimized trajectory parameters. Within the first optimization interval (0-2.5s), the maximum curvature of the original trajectory is 0.45, which is reduced to 0.32 after processing with the variable structure filter, while maintaining the trajectory deviation within 0.15mm.

[0185] To evaluate the quality of the optimized trajectory, the integral value of acceleration S = ∫|a(t)|dt and the maximum trajectory deviation D = max|Poptimized-Poriginal| are calculated and compared with the preset parameter thresholds Smax and Dmax. If S ≤ Smax and D ≤ Dmax, a smooth trajectory is output; otherwise, the filter parameters are adjusted and re-optimized. Taking machining as an example, the preset Smax = 25 m / s and Dmax = 0.2 mm. After the first optimization, S = 28.3 m / s and D = 0.18 mm are calculated. Although D meets the requirements, S exceeds the limit. Therefore, the filter parameters are adjusted, the position weight is reduced to 0.5, and the velocity weight is increased to 0.4. After recalculation, S = 23.6 m / s and D = 0.19 mm are obtained, which meets the preset requirements, and the optimized trajectory is output.

[0186] In practical applications, a CNC machine tool performs machining of a complex contour. The initial planned trajectory contains 300 points. After processing using the method described above, the acceleration variation of the final trajectory is reduced by 47%, and the maximum jerk decreases from 12 m / s³ to 6.3 m / s³, while maintaining a trajectory accuracy deviation of no more than 0.18 mm, meeting the machining accuracy requirements. Compared with traditional fixed-parameter filtering, the trajectory processed using the piecewise adaptive optimization model maintains higher accuracy at corners and achieves better speed stability on straight sections, resulting in an overall performance improvement of 35%.

[0187] Figure 5 Here is a flowchart of the adaptive piecewise filtering process for trajectory optimization in an embodiment of the present invention:

[0188] This flowchart illustrates a complete trajectory optimization process. It begins with the initial trajectory, calculating its curvature, acceleration, and process parameters, and constructing a comprehensive evaluation index based on these parameters, while simultaneously establishing an adaptive threshold function. Next, the system divides the trajectory into optimization intervals according to the adaptive threshold function, determining the boundary points of each interval. Based on this, a variable structure filter containing a state transition matrix and a state vector is established, which adaptively adjusts the state estimation. The system then calculates the state estimates and uses them as the optimized trajectory parameters, while also calculating the integral value of acceleration and the maximum trajectory deviation. The final stage of the process incorporates a judgment mechanism: if the preset parameter thresholds are met, a smooth trajectory is output as the final result; otherwise, the process returns to adjust the filter parameters and recalculate the state estimation, forming a closed-loop optimization process until a satisfactory optimized trajectory is obtained. This adaptive segmented optimization method effectively handles discontinuities in the trajectory, ensuring the smoothness and continuity of the optimization results.

[0189] A second aspect of this invention provides a real-time optimization system for welding trajectories of industrial robots based on deep learning, comprising:

[0190] The first unit is used to collect three-dimensional point cloud data of complex curved surfaces, and to perform noise reduction and feature enhancement on the three-dimensional point cloud data using a multi-scale decomposition algorithm, and to generate high-precision curved surface data by combining a curvature-constrained mesh reconstruction method.

[0191] The second unit is used to establish a dynamic feature space on the high-precision surface data, extract the weld trajectory point set in the dynamic feature space, and construct a weld feature descriptor that integrates geometric features and topological relationships.

[0192] The third unit is used to take the welding current, voltage and wire feed speed as joint optimization objectives based on the weld feature descriptor, and to calculate the optimal combination of process parameters and welding posture sequence through multi-objective reward function iterative calculation, and to establish the mapping relationship between the optimal combination of process parameters and the welding posture sequence.

[0193] The fourth unit is used to map the mapping relationship to the robot joint space, perform obstacle avoidance verification using an improved fast collision detection algorithm, and generate the initial motion trajectory in combination with dynamic constraints.

[0194] The fifth unit is used to establish a segmented adaptive optimization model for the initial motion trajectory. It divides the optimization interval according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory.

[0195] The sixth unit is used to convert the optimized initial motion trajectory into robot control commands to achieve adaptive adjustment of the welding process.

[0196] A third aspect of the present invention provides an electronic device, comprising:

[0197] processor;

[0198] Memory used to store processor-executable instructions;

[0199] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0200] 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.

[0201] 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.

[0202] 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. A real-time optimization method for welding trajectories of industrial robots based on deep learning, characterized in that, include: Three-dimensional point cloud data of complex curved surfaces are collected, and the three-dimensional point cloud data are denoised and feature-enhanced using a multi-scale decomposition algorithm. High-precision curved surface data is generated by combining the curvature-constrained mesh reconstruction method. A dynamic feature space is established on the high-precision surface data, and the weld trajectory point set in the dynamic feature space is extracted to construct a weld feature descriptor that integrates geometric features and topological relationships. Based on the weld feature descriptor, welding current, voltage, and wire feed speed are used as joint optimization objectives. The optimal combination of process parameters and welding posture sequence are calculated iteratively through a multi-objective reward function. A mapping relationship between the optimal combination of process parameters and the welding posture sequence is established, including: Based on the weld feature descriptor, the range of process parameter variation is determined, and process constraints between welding current, welding voltage, and wire feed speed are established. The welding current, welding voltage, and wire feed speed are then constructed as a joint optimization objective. The penetration depth control factor is calculated based on the welding current, the welding voltage and the wire feeding speed, and the weld formation index is calculated based on the geometric parameters of the weld cross-section. The penetration depth control factor and the weld formation index are then used to construct a process quality evaluation index. Collect welding torch posture angle and robot joint position data, calculate the deviation value of the welding torch posture angle relative to the target posture, calculate the motion smoothness of the robot joint position data, and construct the deviation value and the motion smoothness as motion performance evaluation index. The process quality evaluation index and the motion performance evaluation index are combined into a multi-objective reward function. The joint optimization objective is iteratively optimized based on the gradient information of the multi-objective reward function. When the optimization gradient is less than a preset gradient threshold, the optimal combination of process parameters is output. The welding posture sequence is optimized based on the optimal combination of process parameters, and a mapping relationship is established between the weld feature descriptor and the optimal combination of process parameters and the welding posture sequence to achieve collaborative planning of process parameters and welding posture. The mapping relationship is mapped to the robot joint space, and an improved fast collision detection algorithm is used for obstacle avoidance verification. The initial motion trajectory is generated in combination with dynamic constraints. A segmented adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory. The optimized initial motion trajectory is converted into robot control commands to achieve adaptive adjustment of the welding process.

2. The method according to claim 1, characterized in that, The 3D point cloud data is denoised and feature-enhanced using a multi-scale decomposition algorithm, and high-precision surface data is generated by combining a curvature-constrained mesh reconstruction method. The three-dimensional point cloud data is decomposed into multiple scales based on wavelet transform, and the local curvature change rate at each decomposition scale is calculated. The local curvature change rate is then used to perform multi-scale noise reduction on the three-dimensional point cloud data to obtain the noise-reduced point cloud data. Calculate the local neighborhood covariance matrix for the denoised point cloud data, extract the eigenvalues ​​and eigenvectors of the local neighborhood covariance matrix, and construct a local geometric feature descriptor containing curvature information and normal vector information based on the eigenvalues ​​and eigenvectors. The feature distance between the local geometric feature descriptors is calculated using the Gaussian kernel function. Based on the feature distance and the local point cloud density distribution, feature enhancement and density equalization are performed to generate feature-enhanced point cloud data. A local curvature tensor field is constructed on the feature-enhanced point cloud data. The local curvature change is estimated based on the local curvature tensor field. The local curvature change is transformed into a curvature constraint condition. The sampling strategy is determined in combination with the point cloud density to generate a shape quality index that satisfies the local curvature change. The shape quality index, the curvature constraint, and the point cloud fitting error are iteratively optimized and adjusted to adjust the mesh structure until convergence, at which point high-precision surface data is output.

3. The method according to claim 2, characterized in that, A local curvature tensor field is constructed on the feature-enhanced point cloud data. Local curvature changes are estimated based on the local curvature tensor field. These local curvature changes are then transformed into curvature constraints. A sampling strategy is determined by combining the point cloud density, and shape quality indices that satisfy the local curvature changes are generated, including: The local curvature tensor field contains curvature components of each point in the three coordinate axes. Based on the curvature components, eigenvalue equations are constructed. The principal curvature values ​​are obtained by solving the equations whose determinant is zero. The Gaussian curvature is obtained by multiplying the principal curvature values, and the average curvature is obtained by calculating the arithmetic mean of the principal curvature values. The curvature gradient vector is obtained by calculating the partial derivatives of the Gaussian curvature and the average curvature in the three coordinate axis directions. The local curvature change rate is obtained by taking the square root of the sum of the squares of the components of the curvature gradient vector. The local curvature change rate is compared with a preset curvature threshold, and the constraint coefficient is determined based on the comparison result. When the local curvature change rate meets the preset curvature threshold, the local curvature change rate is subtracted from the preset curvature threshold and then divided by a smoothing factor. The opposite number is used as the power of the exponential function to calculate the constraint coefficient. Based on the constraint coefficients, a spherical search region is constructed in the point cloud data with each data point as the center. The number of points in the spherical search region is divided by the area of ​​the spherical search region to obtain the point density. The ratio of the standard deviation to the mean of the point density is calculated to obtain the shape quality index.

4. The method according to claim 1, characterized in that, Calculate the deviation of the welding torch posture angle relative to the target posture, calculate the motion smoothness of the robot joint position, and construct a motion performance evaluation index from the deviation value and the motion smoothness, including: Calculate the angular deviation between the welding torch attitude angle and the target attitude angle, multiply the angular deviation by the corresponding direction weight coefficient and sum them to obtain a weighted deviation index, which reflects the importance of different directional deviations; The position and velocity signals of each joint of the robot are collected. The second derivative of the position signal is obtained to obtain the velocity difference signal. The absolute value of the velocity difference signal is calculated to obtain the position fluctuation index, which characterizes the acceleration change of the joint movement. The velocity difference increment signal is obtained by taking the first derivative of the velocity signal, and the velocity change rate is obtained by calculating the absolute value of the velocity difference increment signal. The position fluctuation index and the velocity change rate are combined to obtain the motion stability index. Based on the weighted deviation index and the motion stability index, a motion performance evaluation index is constructed using performance weighting coefficients, which are used to balance the influence of attitude accuracy and motion stability.

5. The method according to claim 1, characterized in that, The mapping relationship is mapped to the robot joint space, and an improved fast collision detection algorithm is used for obstacle avoidance verification. The initial motion trajectory is generated by combining dynamic constraints, including: The mapping relationships in the workspace are transformed to the robot joint space through inverse kinematics to obtain the joint angle sequence; An improved fast collision detection algorithm is used for obstacle avoidance verification. The improved fast collision detection algorithm includes: constructing a hierarchical axial bounding box tree structure based on the joint angle sequence, updating the spatial position parameters of the axial bounding box tree structure in real time as the joint angle sequence changes; calculating the minimum distance between adjacent bounding boxes in the axial bounding box tree structure, comparing the minimum distance with a preset safety threshold, and marking a collision risk area when the minimum distance is less than the preset safety threshold. A robot dynamics model is established, which includes a joint inertia matrix, a velocity coupling matrix, and a gravity term. Based on the robot dynamics model, position limits, velocity limits, acceleration limits, and driving torque limits of joint motion are set. The position limits, velocity limits, acceleration limits, and driving torque limits are used as constraints. An initial motion trajectory that satisfies the collision risk region is generated according to the constraints.

6. The method according to claim 1, characterized in that, A piecewise adaptive optimization model is established for the initial motion trajectory. The optimization interval is divided based on a comprehensive evaluation index of trajectory curvature, acceleration rate of change, and process requirements. Within the optimization interval, a variable structure filtering algorithm is used for trajectory smoothing. This dynamic optimization of the initial motion trajectory includes: Calculate the curvature, acceleration, and process parameters of the initial motion trajectory; combine the curvature, acceleration, and process parameters into a comprehensive evaluation index; and construct an adaptive threshold function based on the comprehensive evaluation index. The initial motion trajectory is divided into multiple optimization intervals based on the value of the adaptive threshold function. The adaptive threshold function decays exponentially with the change of the comprehensive evaluation index. The time point when the comprehensive evaluation index is greater than the value of the adaptive threshold function is determined as the interval boundary point. A variable structure filter is established within the optimization interval based on the boundary points of the interval. The variable structure filter includes a state transition matrix and a state vector. A state estimate is calculated based on the observed values ​​of the state vector, and the state estimate is used as the optimized trajectory parameters. Calculate the integral value of acceleration and the maximum deviation of trajectory parameters after optimization. When the integral value of acceleration and the maximum deviation of trajectory meet the preset parameter threshold, output a smooth trajectory. When they do not meet the preset parameter threshold, return to adjust the parameters of the variable structure filter until an optimized trajectory that meets the smoothness requirement is generated.

7. A real-time optimization system for welding trajectories of industrial robots based on deep learning, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect three-dimensional point cloud data of complex curved surfaces, and to perform noise reduction and feature enhancement on the three-dimensional point cloud data using a multi-scale decomposition algorithm, and to generate high-precision curved surface data by combining a curvature-constrained mesh reconstruction method. The second unit is used to establish a dynamic feature space on the high-precision surface data, extract the weld trajectory point set in the dynamic feature space, and construct a weld feature descriptor that integrates geometric features and topological relationships. The third unit is used to take the welding current, voltage and wire feed speed as joint optimization objectives based on the weld feature descriptor, and to calculate the optimal combination of process parameters and welding posture sequence through multi-objective reward function iteration, and to establish the mapping relationship between the optimal combination of process parameters and the welding posture sequence. The fourth unit is used to map the mapping relationship to the robot joint space, perform obstacle avoidance verification using an improved fast collision detection algorithm, and generate the initial motion trajectory in combination with dynamic constraints. The fifth unit is used to establish a segmented adaptive optimization model for the initial motion trajectory. It divides the optimization interval according to the comprehensive evaluation index of trajectory curvature, acceleration change rate and process requirements. Within the optimization interval, a variable structure filtering algorithm is used to smooth the trajectory, thereby realizing the dynamic optimization of the initial motion trajectory. The sixth unit is used to convert the optimized initial motion trajectory into robot control commands to achieve adaptive adjustment of the welding process.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images