Feature extraction method for gusset point cloud based on bidirectional euclidean constraint and quartile distance statistical truncation

Abstract

The application relates to the fields of machine vision and intelligent welding, and discloses a fillet weld point cloud feature extraction method based on bidirectional Euclidean constraint and quartile distance statistical truncation. In non-standard flexible sheet processing, the theoretical features are suspended due to thermal deformation, the side wall topological features are misdirected, and the end points of the teaching-free system are deviated due to point solid welding splashing, pose singularity and interference collision and other composite pain points; the application is characterized in that the space affine dimension reduction is realized through the Rodriguez transformation, the side wall interference is accurately removed through the rotating minimum circumscribed rectangle feature, the entity point array deeply embedded in the gap is extracted through the bidirectional Euclidean projection joint evaluation model, the random assembly gap is resisted through the multiplicative feedback state machine, the non-parametric statistical truncation mechanism of the interquartile range (IQR) is introduced after the fitting feature, the heavy-tailed noise is adaptively stripped to accurately position the physical end point, the TCP pose is reconstructed through the Schmidt orthogonalization, the four-point anti-collision servo track is generated, and high-fidelity dynamic tracking and safe anti-collision guidance under extreme working conditions are realized.

Description

Technical Field

[0001] This invention relates to the fields of machine vision, point cloud processing and intelligent flexible manufacturing technology, and in particular to a method for extracting fillet weld point cloud features based on bidirectional joint Euclidean constraints and nonparametric quantile truncation based on interquartile range (IQR) under strong noise resistance. Background Technology

[0002] In the field of small-batch flexible manufacturing of cabinet sheet metal, automotive parts and non-standard parts, "teaching-free" automated robotic welding and real-time weld seam tracking are the core development directions of the industry. In such scenarios, traditional preset trajectory programs are completely unable to adapt to the variability of non-standard parts in size and the randomness of assembly. It is necessary to rely on 3D vision systems to extract weld seam point cloud features in real time to automatically plan the welding trajectory.

[0003] However, existing point cloud feature extraction algorithms face significant technical bottlenecks in the flexible manufacturing of thin plates and non-standard parts:

[0004] First, topological misleading effects on the sidewalls of thin plates: when dealing with thin plates (such as those with a thickness of...) When the steel plate is overlapped or corner-jointed, the existing three-dimensional plane segmentation algorithm is very likely to misidentify the thickness sidewall of the steel plate as a "valid working master plane", which will cause the subsequent plane feature matching logic to completely collapse.

[0005] Second, thermal deformation and gaps cause theoretical characteristics to be suspended: thin plates are prone to severe thermal warping deformation during welding. Coupled with the inherent blanking tolerances of non-standard parts, unpredictable dynamic gaps exist at the root of the actual fillet weld. The "pure mathematical intersection line" obtained by the traditional algorithm by fitting the plane will be seriously suspended outside the actual metal surface. For the root pass welding condition with large gaps, it is very easy to cause missed welds. In addition, pure Euclidean distance often fails to address the anisotropic noise scattering caused by high reflectivity.

[0006] Third, the endpoints are susceptible to spatter from spot welding: When cutting the start and end points of the weld to plan the path, spot welding spatter or cutting burrs often appear at the corners of thin plate splicing. Using the traditional absolute extreme value method may cause catastrophic distortion of the extracted endpoint coordinates.

[0007] Fourth, Jacobi singularity in teachless attitude calculation: In a teachless system, extracting only a straight weld line is not enough to guide the robot's operation. Traditional normal average attitude calculation methods lack strict algebraic orthogonalization constraints, and the resulting attitude matrix is ​​prone to slight distortion. When guiding the welding torch into an unknown angle, it is very easy to cause "singularity" errors in the robot arm's inverse kinematics and high-frequency joint vibrations, and it does not guarantee uniform heat input and penetration of the thin plates on both sides. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for extracting fillet weld point cloud features based on bidirectional Euclidean constraints and interquartile distance statistical truncation. This method addresses the combined challenges of sidewall interference in thin plate deformation, suspended gaps in non-standard component assembly, and singular attitudes without teaching from the algebraic geometric level, thereby achieving high-precision dynamic tracking and trajectory generation under extreme flexible working conditions.

[0009] To achieve the above objectives, the present invention includes the following core technical solutions:

[0010] First, a Rodrigues rotation formula transformation is introduced to construct a spatial affine matrix, forcibly reducing the dimensionality of the 3D point cloud topological features to the 2D image domain. The short side of the 2D minimum bounding rectangle is extracted to adaptively filter out interference from the thin plate's thickness sidewalls. Second, abandoning the "intersection of simultaneous equations" logic, a joint evaluation model of bidirectional Euclidean projection distance is innovatively constructed. This generates a "controlled rhombic prism search envelope" in physical space, directly extracting the real metal scattered points deeply embedded in the assembly gaps. Subsequently, a controlled closed-loop multiplicative feedback state machine is designed to dynamically expand the envelope threshold and counteract interference from the thin plate's thickness sidewalls in real time. For dynamic random gaps caused by thermal deformation of thin plates or non-standard parts cutting; after extracting physical straight line features, a non-parametric statistical truncation mechanism based on interquartile range (IQR) is introduced to adaptively remove outliers at both ends of the sequence, replacing the fragile absolute extremum method, robustly removing heavy-tailed noise at both ends to accurately lock the start and end point features; finally, based on the Schmidt orthogonalization principle, the absolutely orthogonal TCP pose matrix is ​​reconstructed, and the absolutely orthogonal precise X-axis vector, Y-axis vector and main proximity Z-axis vector are extracted to construct a homogeneous rotation transformation matrix, which is then reverse-analyzed into the robot's Euler angle pose.

[0011] The present invention has the following significant beneficial effects:

[0012] 1. Completely eliminate interference from the sidewalls of thin plates and thermal deformation distortion: This invention eliminates the logical misleading of the thickness surface of thin plates through dimensional reduction; at the same time, it uses a two-way Euclidean constraint evaluation model to convert geometric intersection into physical lattice interception, which includes the dynamic gap caused by the thermal warping deformation of the thin plate in real time, perfectly adapting to the teaching-free flexible processing of non-standard parts.

[0013] 2. Extremely strong noise resistance of the teaching-free endpoint features: In response to the heavy-tail noise caused by corner welding spatter, this invention uses a non-parametric statistical truncation mechanism based on interquartile range (IQR) to adaptively remove outliers at both ends of the sequence. By relying on non-parametric statistical laws, it accurately removes the heavy-tailed distribution of spot welding spatter noise at the corner, ensuring the absolute accuracy of the arc initiation point coordinates.

[0014] 3. Natively eliminates singularities in the inverse kinematics of the robotic arm and ensures uniform penetration: This invention forcibly introduces double outer product orthogonalization constraints in the no-teach posture generation stage to ensure the strict orthogonality of the homogeneous transformation matrix corresponding to the feature points from the underlying algebraic principle, which can ensure that the robotic arm performs posture interpolation smoothly without jitter; at the same time, it natively ensures that the welding torch always cuts into the corner seam with "absolute symmetrical bisecting angle", maximizing the uniform penetration of the thin plate on both sides. Attached Figure Description

[0015] Figure 1 This is the overall point cloud feature extraction algorithm architecture and data flow control diagram of the present invention.

[0016] Figure 2 This is a measured result of the dual-reference plane segmentation and intersection feature extraction based on the Rodrigues transform space dimensionality reduction and removal of sidewall interference according to the present invention.

[0017] Figure 3 This is a schematic diagram comparing the topology of the weld feature line reconstructed by the algorithm of this invention with the original undivided scattered point cluster.

[0018] Figure 4 This is a magnified view of the local effect of the precise positioning of the physical start and end points based on the bidirectional Euclidean constraint and IQR statistical truncation mechanism of the present invention.

[0019] Figure 5 This is a schematic diagram illustrating the derivation of the robot's six-DOF TCP servo pose and four-point safe trajectory planning based on Schmidt orthogonal reconstruction according to the present invention. Detailed implementation method.

[0020] To make the objectives and underlying mathematical principles of this invention clearer, the following detailed description of each core algorithm module of this invention is provided in conjunction with examples of non-standard parts and thin-plate automated robots without instruction.

[0021] Step 1: Spatial Dimensionality Reduction Evaluation and Benchmark Surface Feature Locking to Combat Topological Distortion

[0022] After extracting multiple candidate planes, in order to peel them off... The inherent thickness of the thin plate sidewalls leads to topological misleading of spatial pairing logic. In the following embodiments, a three-dimensional dimensionality-reduced affine evaluation model is constructed.

[0023] Extract the normalized unit normal vector of the features of any candidate plane. Calculate its basis vector relative to the Z-axis of the absolute coordinate system. spatial angle .

[0024] Extracting the rotation axis vectors describing the tilt attitude in planar space :

[0025]

[0026] Based on Rodrigues' rotation theorem, construct the rotation axis vector. antisymmetric matrix Analytically solve for the three-dimensional affine rotation matrix that forces it to flatten into a two-dimensional plane. :

[0027]

[0028]

[0029] Apply a rotation transformation matrix to the three-dimensional point set of the candidate plane. The Z-axis component is removed, and a two-dimensional pixel set is generated by dimensionality reduction. The minimum bounding rectangle of this two-dimensional pixel set is calculated. If its shortest side (i.e., the thickness mapping dimension) is less than a preset physical plate thickness threshold, the plane is marked as a sidewall interference surface and permanently removed. The retained main working surface is then identified as the first welding reference surface. With the second welding reference surface .

[0030] Specifically, from Figure 2 and 3 It can be seen that after spatial dimensionality reduction evaluation and segmentation, the system can extract data from the original scattered point cloud background (e.g., ...). Figure 3 The first reference plane, which is a spliced ​​structure, is obtained by precisely peeling away the gray background point cloud shown. Figure 2 Blue point cloud) and the second reference plane ( Figure 2 Green point cloud), preliminary fitting yields characteristic intersection lines that fit the physical root ( Figure 2 The red linear feature enables effective topological decoupling of interference from the thin plate sidewalls.

[0031] Steps 2-3: Construct bidirectional Euclidean constraints to correct deformation deviations in real time.

[0032] right and The plane equations are normalized to obtain the coefficient matrix. and (Strictly meet) Set the initial physical space tolerance threshold. .

[0033] Construct a joint projection distance evaluation function for point cloud space. any discrete scatter point in Only if this point strictly satisfies:

[0034]

[0035] and

[0036]

[0037] At that time, it was included in the feature point set of the physical transition zone of the fillet weld in round 0. middle.

[0038] This constraint forcibly filters out the large number of redundant points on the flat wall surface, accurately captures the characteristic scattered points of the metal entity deep in the hinge assembly gap, and is not affected by the overall thermal warping deformation of the thin plate or the macroscopic influence of non-standard cutting tolerances, laying a solid foundation for dynamic tracking.

[0039] Step 4: Controlled closed-loop multiplicative feedback state machine to resist random backlashes in non-standard components

[0040] In non-standard flexible manufacturing, the weld gap width changes constantly. This invention designs an adaptive state machine and sets a safe lower limit for high-fidelity fitting feature points. With the coefficient of state expansion (like ).

[0041] In the After the first iteration of the truncation, if the cardinality of the feature point set is found... The system automatically triggers multiplicative feedback, forcibly expanding the search envelope boundary:

[0042]

[0043] Utilizing the new tolerance threshold of expansion amplification By reconstructing a larger rhombic prism envelope for feature extraction and iterating in a closed loop until enough high-fidelity transition points are obtained, the limitations of rigid pre-set procedures are completely eliminated.

[0044] Steps 5-6: High-fidelity fitting and IQR-based nonparametric statistical truncation

[0045] For the feature point set of the transition zone, robust linear regression with Huber loss function is used for fitting to reconstruct the true geometric straight line features of the fillet weld. .

[0046] Introducing nonparametric quantile extraction endpoint anti-skewing: Orthogonally projecting the extracted planar physical outer edge points onto the straight line feature. Obtain a one-dimensional relative projection coordinate sequence To address the issue of solder spatter noise at corner points, this invention abandons the traditional extreme value method and fixed proportion quantile method, and uses the statistical interquartile range (IQR) truncation algorithm to calculate the 25th percentile. and 75th percentile :

[0047]

[0048]

[0049] Calculate interquartile range And construct confidence boundaries for the distribution characteristics of heavy-tailed noise caused by workpiece corner chamfering and spot welding spatter:

[0050]

[0051] The farthest point within the confidence boundary is retained as the starting point of the arc. and the point of convergence This perfectly filters out isolated splash points that deviate from the main distribution, ensuring absolute accuracy of the arc initiation and termination coordinates. The measured effect of this endpoint locking logic is as follows: Figure 4 As shown, by introducing the interquartile range (IQR) statistical confidence boundary, outlier noise (i.e., spot welding spatter noise outside the direction indicated by the red arrow in the figure) distributed outside the actual physical boundary can be adaptively identified and shielded, so that the endpoint coordinates are strictly locked at the physical edge position of the metal body, and the endpoint pull error is completely eliminated.

[0052] Step 7: Schmidt Orthogonalization TCP Reconstruction and Safe Trajectory Planning:

[0053] Get and Polarity reversed and the principal normal vectors that are uniformly opposite to the workpiece entity and .

[0054] Solve for the robot's principal approach Z-axis vector that precisely bisects the unknown spatial angle to ensure uniform heat input to the two thin plates:

[0055]

[0056] Using the direction vector of the geometric straight line as the initial temporary axis Force the execution of the quadratic outer product operation of the Schmitt orthogonalization variant:

[0057] Calculate the precise Y-axis perpendicular to the action plane:

[0058]

[0059] Reverse calculation of strictly orthogonal precise X-axis:

[0060]

[0061] Set the space safety lift offset scalar Mathematically, a reverse rigid body translation is performed along the normal Z-axis on the starting and ending points:

[0062]

[0063]

[0064] The final output consists of precisely orthogonal column vectors. This constitutes a homogeneous rotation transformation matrix. The homogeneous transformation matrix is ​​losslessly analyzed into Euler angle commands for the robot, combined with safe entry. The teach-free, collision-avoiding, safe, continuous servo trajectory not only mathematically prevents universal joint deadlock but also perfectly avoids collision interference caused by the welding torch blindly entering or exiting through geometric offset.

[0065] The final generated job trajectory and pose are as follows Figure 5 As shown, the RGB three-dimensional coordinate system markers generated at both ends of the weld are absolutely orthogonal TCP poses reconstructed by Schmidt orthogonalization. Combined with the safety offset to generate a four-point trajectory sequence, as shown in the yellow / red continuous path in the figure, this invention completes the kinematic closed loop from high-fidelity extraction of visual features to collision-free execution of the robot.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting feature data from fillet weld point clouds based on bidirectional Euclidean constraints and interquartile range statistical truncation, characterized in that, Includes the following steps: Step 1: Obtain a three-dimensional scattered point cloud of the fillet weld containing the two-sided splicing structure, perform a spatial affine transformation evaluation of the three-dimensional point cloud, and extract and lock the first welding reference plane and the second welding reference plane of the fillet weld. Step 2: Establish a joint evaluation model of spatial features based on bidirectional Euclidean projection distance, and set an initial spatial tolerance threshold; Step 3: Traverse the 3D scattered point cloud and calculate the Euclidean projection distance of each spatial point to the first welding reference plane and the second welding reference plane respectively; forcibly extract spatial points whose bidirectional projection distance is strictly less than the current spatial tolerance threshold, and construct a physical transition zone point cloud set; Step 4: Construct a controlled closed-loop multiplicative feedback state machine, count the feature point base of the physical transition zone point cloud, and when the base is less than the set high-fidelity fitting safety lower limit, multiplicatively amplify the current spatial tolerance threshold according to the preset expansion coefficient, and reconstruct the physical transition zone point cloud using the updated spatial tolerance threshold. Iterate in the closed loop until the feature point base meets the requirements. Step 5: Perform spatial linear regression fitting on the qualified physical transition zone point set to extract the true geometric linear features of the fillet weld. Step 6: Extract the set of physical outer edge points of the plane and project them onto the geometric straight line features of the real fillet weld to generate a one-dimensional relative projection coordinate sequence and perform monotonically sorting. Introduce a non-parametric statistical truncation mechanism based on interquartile range (IQR) to adaptively remove outliers at both ends of the sequence and accurately locate the absolute physical start and end point features of the fillet weld. Step 7: Then extract the normal vectors of the first and second welding reference planes, and combine them with the geometric straight line features of the real fillet weld to perform spatial vector double outer product orthogonalization reconstruction based on Gram-Schmidt Orthogonalization, and output the absolutely orthogonal pose of the robot's six-degree-of-freedom servo tool center point (TCP).

2. The method according to claim 1, characterized in that, The specific mathematical analysis process for performing the spatial affine transformation evaluation of the 3D point cloud in step 1 is as follows: Extracting the unit normal vector of candidate planes from point cloud data Calculate the unit normal vector and the basis vector of the global coordinate system Z-axis. spatial angle and rotation axis vector ; Construct rotation axis vector antisymmetric matrix ; Construct a three-dimensional homogeneous rotation matrix based on Rodrigues' Rotation Formula. ; in, Rotation axis vector antisymmetric matrix, It is the identity matrix; Using the rotation matrix The candidate planar 3D point set is forcibly reduced in dimension and projected onto a 2D reference plane. The minimum bounding rectangle of its 2D pixel set is calculated. The shortest side dimension of this rectangle is compared with a preset physical thickness threshold for thin plates. Sidewall thickness interference surfaces with short side dimensions smaller than the threshold are eliminated, and the remaining main working surface is the first welding reference plane. With the second welding reference surface .

3. The method according to claim 1, characterized in that, The formulas for constructing the joint evaluation model of bidirectional Euclidean space features and extracting the point cloud set of the physical transition zone in steps 2 and 3 are as follows: Let the first welding reference surface With the second welding reference surface The coefficient matrices of the four-dimensional plane equations are respectively and And it satisfies the unitization condition. ; For any point in the 3D point cloud data Its distance to the reference plane Euclidean projection distance The calculation formula is ; Construct joint equations with two-way physical constraints, and extract lattice points that satisfy the conditions to form the first... Physical transition zone points of each round In the formula, For the first The current spatial tolerance threshold for the next iteration.

4. The method according to claim 1, characterized in that, The adaptive update logic of the controlled closed-loop multiplicative feedback state machine in step 4 is as follows: Let the lower limit of safety for feature points be... The set state expansion coefficient is When physical transition zone points gather Feature point cardinality Trigger the multiplicative feedback mechanism to update the spatial tolerance threshold for the next round. ; Clear the current point set cache and utilize the updated magnified value. Return to reconstruct a new physical transition region point cloud until the feature point cardinality is reached. .

5. The method according to claim 1, characterized in that, The specific mathematical implementation of the quantile truncation mechanism based on nonparametric statistics of the interquartile range (IQR) introduced in step 6 is as follows: Let the extracted true fillet weld geometric straight line features be... The extracted planar physical outer edge point lattice is orthogonally projected onto the straight line. Calculate the one-dimensional topological coordinates of each projection point relative to the reference origin along the straight line extension direction, thus forming a one-dimensional relative projection coordinate sequence. ; and sequence Sort monotonically by numerical value; The distribution characteristics of heavy-tail noise caused by chamfering of workpiece corners and spatter from spot welding were calculated. First quartile With the third quartile ; Calculate interquartile range And define the statistically effective physical cutoff interval as ; Extract the extreme values ​​of the sequence boundary coordinates that fall within the effective physical cutoff interval, and use them as the absolute physical start and end cutoff points of the fillet weld to eliminate endpoint pull-off errors.

6. The method according to claim 1, characterized in that, The specific calculation process for reconstructing the spatial vector outer product orthogonality in step 7 is as follows: Based on the line connecting the centroids of the two reference planes in physical space, the polarity of the normal vector is corrected by performing an inner product sign adjustment to obtain a principal normal vector pointing uniformly to the outside of the included angle of the fillet weld. and ; By summing the normal vectors of uniform polarity and then normalizing them in reverse, a robot principal approach Z-axis basis vector that rigorously bisects the spatial angle is generated. ; The unit extension vector of the geometric straight line feature of the actual fillet weld is used as the initial X-axis. The precise Y-axis in the TCP servo guidance pose matrix is ​​constructed using double cross product reconstruction based on Gram-Schmidt Orthogonalization. and precise X-axis as ; The above-mentioned start and end cutoff points are defined as the arc starting points. With the end of the arc Set the space safety lifting bias scalar Along the main approach Z-axis vector Apply an offset translation in the opposite direction to reconstruct the safe entry point. and safe departure point ; Finally, from the exact orthogonal column vectors Construct homogeneous rotation transformation matrices, forming the... The homogeneous transformation matrix is ​​losslessly parsed into Euler angle commands for the robot, combined with... This constitutes a teaching-free, collision-avoidance, safe, continuous servo trajectory.

Images