Automobile parts assembly control method based on visual positioning feedback

Abstract

The application relates to the technical field of machine vision, and discloses a kind of automobile parts assembly control method based on visual positioning feedback, comprising: obtaining glue groove and transparent lampshade image and establishing circumferential coordinate system;Extract the refraction phase field generated when the transparent lampshade approaches the glue groove to peel off the refraction artifact interference;Construct the particle state containing pose deviation, edge deformation and glue strip deformation variable, generate the outer edge deformation line using the state;Based on the refraction phase field, deformation cost and volume cost, an adaptive function is established;Perform adaptive particle swarm optimization search to output the optimal particle, calculate the optimization direction to generate the pose correction amount to adjust the mechanical arm posture;Finally, based on the closure cost and volume cost, the final pose is obtained, the robot is controlled to be in place along the continuous pressing path and the visual record is generated.The application suppresses the positioning deviation caused by refraction artifact and improves the assembly precision.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and more specifically, to a method for controlling the assembly of automotive parts based on visual positioning feedback. Background Technology

[0002] In the automated assembly process of automotive headlights, pressing the transparent polycarbonate (PC) outer lens into the peripheral adhesive groove of the black polypropylene (PP) housing is a key step. To ensure watertightness and consistency in appearance gaps, sealant is usually pre-applied to the adhesive groove, and then a robot picks up the lens to complete the pressing. In recent years, machine vision systems have been widely introduced in the industry to guide the robot's assembly trajectory.

[0003] However, during actual production line debugging, we found that traditional visual guidance solutions are prone to failure when handling such transparent components. The lampshade itself is a complex curved transparent material. As it approaches the black adhesive groove filled with sealant, the reflections from the edge of the housing, the internal reflector bowl, and the surface of the adhesive strip cause severe light distortion and refraction through the transparent curved surface. This optical phenomenon makes the actual boundary of the adhesive groove and the outer edge of the lampshade appear blurred in the image captured by the camera, forming a false outline that shifts with the assembly process.

[0004] Most existing machine vision algorithms on production lines rely on the extraction and matching of individual edge pixels. When the vision system misjudges these wandering refraction artifacts as real contact boundaries, the output positioning coordinates will have significant deviations. Once this deviation is transmitted to the robot actuator, it will cause circumferential local misalignment when the lamp cover is pressed into the glue groove. This can range from insufficient pressure on one side of the glue strip and excessive compression on the other side causing glue overflow and disrupting the seal continuity, to localized stress concentration between components and even edge warping. Simply increasing camera resolution or adding light sources is insufficient to eliminate artifact interference caused by the material's inherent light transmittance, and existing control logic often only focuses on the alignment of rigid body spatial coordinates, failing to comprehensively consider the minute deformation of transparent components under pressure and the changes in the volume of the glue groove. This results in the current automotive lamp pressing process being highly dependent on manual intervention, making it difficult to achieve high-yield automated control. Summary of the Invention

[0005] This invention provides a method for controlling the assembly of automotive parts based on visual positioning feedback, which solves the technical problems mentioned in the background art.

[0006] This invention provides a vision-based positioning feedback-based method for assembling automotive parts, applied to a robotic assembly system including a camera, for pressing a polycarbonate transparent lampshade into a peripheral adhesive groove of a polypropylene housing, comprising: Images of the periphery of the polypropylene shell and the polycarbonate transparent lampshade are obtained, and a circumferential coordinate system is established and bound to the robot assembly system to unify the normal reference of the lampshade and the glue groove. In the circumferential coordinate system, the refracted phase field generated when the polycarbonate transparent lampshade approaches the peripheral adhesive groove is extracted to remove the refraction artifact interference caused by the transparent curved surface; Construct a particle state that characterizes assembly bias and elastic deformation. The particle state includes pose deviation, edge deformation amount and rubber strip shape change. The particle state is used to generate an outer edge deformation line that characterizes the actual contact contour. Based on the refracted phase field, a fitness function is established by combining the deformation cost characterizing physical stress and the volume cost characterizing adhesive overflow. Perform adaptive particle swarm visual feedback search and output the optimal particle that minimizes the fitness function; The optimal direction is calculated based on the optimal particle, the pose correction amount is generated, and the robot assembly system is controlled to adjust the spatial posture. The final pose is obtained based on the closure cost and volume cost representing the assembly gap. The robot assembly system is then controlled to complete the positioning along the continuous pressing path, and a visual record is formed simultaneously.

[0007] The beneficial effects of this invention are as follows: Addressing the visual deviations that easily occur during the assembly of polycarbonate transparent lamp covers and surrounding adhesive grooves, this invention no longer relies on edge contour points susceptible to optical interference. Instead, it extracts the refracted phase field generated when the transparent component approaches the adhesive groove and constructs a particle state including pose deviation, edge deformation, and adhesive strip shape variation for adaptive optimization calculation. By integrating visual, deformation, and volume costs, the robot is guided to adjust along a trajectory that avoids local adhesive overflow and component interference, effectively reducing the risk of misalignment during pressing caused by the refraction of the transparent component's curved surface, and improving the sealing fit and gap uniformity of the automotive headlight adhesive groove assembly. Attached Figure Description

[0008] Figure 1 This is a flowchart of an automotive parts assembly control method based on visual positioning feedback according to the present invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] like Figure 1 As shown, a visual positioning feedback-based automotive parts assembly control method is applied to a robot assembly system including a camera, for pressing a polycarbonate transparent lampshade into a peripheral adhesive groove of a polypropylene housing, comprising: Images of the periphery of the polypropylene shell and the polycarbonate transparent lampshade are obtained, and a circumferential coordinate system is established and bound to the robot assembly system to unify the normal reference of the lampshade and the glue groove. In the circumferential coordinate system, the refracted phase field generated when the polycarbonate transparent lampshade approaches the peripheral adhesive groove is extracted to remove the refraction artifact interference caused by the transparent curved surface; Construct a particle state that characterizes assembly bias and elastic deformation. The particle state includes pose deviation, edge deformation amount and rubber strip shape change. The particle state is used to generate an outer edge deformation line that characterizes the actual contact contour. Based on the refracted phase field, a fitness function is established by combining the deformation cost characterizing physical stress and the volume cost characterizing adhesive overflow. Perform adaptive particle swarm visual feedback search and output the optimal particle that minimizes the fitness function; The optimal direction is calculated based on the optimal particle, the pose correction amount is generated, and the robot assembly system is controlled to adjust the spatial posture. The final pose is obtained based on the closure cost and volume cost representing the assembly gap. The robot assembly system is then controlled to complete the positioning along the continuous pressing path, and a visual record is formed simultaneously.

[0011] The S201 acquires the camera's intrinsic parameters and the robot assembly system's extrinsic parameters from the camera. The camera uses an "eye-to-hand" mounting method, fixed 1.5 meters directly above the assembly station. (Camera intrinsic parameter matrix...) Describe the internal optical characteristics of the camera, in the form of ,in and These are the focal lengths along the x and y axes, respectively, in pixels. and Principal point coordinates, in pixels. Camera intrinsic parameters were obtained using the Zhang calibration method. The calibration plate was a 9×7 checkerboard ceramic calibration plate, with each checkerboard square measuring 25 mm × 25 mm. The surface was matte to reduce reflection. Camera extrinsic parameter matrix. Describe the pose relationship between the camera and the base coordinate system of the robot assembly system, where It is a 3×3 rotation matrix. The translation vector is 3×1, and the unit is meters. The camera extrinsic parameters are obtained through hand-eye calibration. During the calibration process, the robot end effector carrying the calibration board moves within the camera's field of view at least 15 different poses, and the extrinsic parameter matrix is ​​calculated after acquiring the corresponding images.

[0012] S202 extracts the center line of the housing glue tank and the outer edge reference line of the lampshade, used to characterize the assembly reference, from the acquired images. First, the acquired images are converted to grayscale, transforming the color images into single-channel grayscale images. Then, the Canny edge detection algorithm is used to extract image edge features. The high and low thresholds of the Canny algorithm are automatically calculated using the Otsu algorithm, with the low threshold being 0.4 times the high threshold. A contour tracking algorithm is used to extract continuous edge contours, with the contour tracking hierarchy rule being to extract the outermost contour and exclude internal holes and background contours. The extracted contours are then Gaussian smoothed with a Gaussian kernel size of 5×5 and a standard deviation of 1.0 to eliminate contour noise. For the housing glue tank, the inner wall contour and outer wall contour are extracted separately, and the center line of the glue tank is calculated using the midpoint method. This involves taking the average coordinates of corresponding points on the inner and outer wall contours. For the lampshade, its outermost contour is extracted as the baseline for the outer edge of the lampshade. .in The normalized circumferential parameter has a value range of [value range missing]. The total arc length of the corresponding contour.

[0013] The grayscale conversion uses a weighted average method, and the calculation formula is as follows: ,in , , These represent the pixel values ​​for the red, green, and blue channels of a color image, respectively, with a range of values ​​ranging from [value range missing]. .

[0014] The formula for calculating the gradient magnitude of the Sobel operator in the Canny edge detection algorithm is as follows: ,in The gradient component is in the x-direction. This represents the gradient component in the y-direction.

[0015] S203 utilizes camera extrinsic and intrinsic parameters to construct a perspective mapping relationship, projecting the target coordinates of the robot assembly system's spatial location onto the camera's two-dimensional plane to eliminate hand-eye coordinate differences. The formula for calculating the perspective mapping relationship is: ; in These are homogeneous pixel coordinates on the camera's two-dimensional plane, in the form of: , and These are pixel coordinate values; These are homogeneous three-dimensional spatial coordinates in the base coordinate system of the robot assembly system, in the form of: , , , These are spatial coordinates, in meters.

[0016] S204 extracts the tangential variation trend of the circumferential parameters of the center line of the glue groove in the shell. After orthogonal rotation and mold length normalization, it generates the glue groove normal vector to indicate the direction of force on the glue groove. First, the center line of the glue groove is... Convert to a 3D coordinate sequence in the robot's base coordinate system Calculate the tangential vector of the glue tank centerline. That is, the three-dimensional coordinates with respect to the normalized circumferential parameters The first derivative of is calculated using the following formula: Obtain the unit vector of the camera's optical axis in the robot's base coordinate system. Its direction is determined by the third column of the rotation matrix of the camera's extrinsic parameters. The tangential vector is converted to the normal direction through a cross product operation, and then normalized to its magnitude to obtain the normal vector of the glue groove. The calculation formula is: ; in Circumferential parameters The three-dimensional unit glue groove normal vector at the location; This represents the cross product operation of three-dimensional vectors. This represents the L2 norm of a three-dimensional vector.

[0017] The unit vector of the camera optical axis in the robot's base coordinate system Extracted from the third column of the camera extrinsic rotation matrix, the calculation formula is as follows: ,in Representing the rotation matrix The third column vector.

[0018] S205 uses circumferential parameters as a unified variable to align the centerline of the housing's glue groove with the outer edge baseline of the lampshade into the same circumferential coordinate system. First, the starting point of the glue groove's centerline is determined by selecting the center of a pre-set injection positioning hole on the glue groove as the starting point, corresponding to the normalized circumferential parameters. Calculate the total arc length of the center line of the glue groove and the baseline of the outer edge of the lampshade, respectively. and The unit is meters. The arc lengths of both contours are normalized to... The interval is used to obtain the normalized circumferential parameters. Mapping relationship with contour points. For any point on the outer edge baseline of the lampshade, find its corresponding normalized circumferential parameter through arc length interpolation. This achieves circumferential alignment with the center line of the glue tank.

[0019] S301 defines the normal scanning area for excluding background interference based on the inner and outer wall boundaries of the surrounding adhesive tank. For circumferential parameters... The glue tank at the location, with the center line of the glue tank Centered on the normal vector of the glue groove The direction extends inwards and outwards to define the normal scan area. The width of the normal scanning area is 1.2 times the actual width of the glue tank, which is obtained from the CAD model and is in meters. Normal scanning area The range is ,in , , Circumferential parameters The width of the glue tank at that location is in meters.

[0020] Within the normal scanning area, S302 extracts the current gradient, representing the current visual feature, and the reference gradient, representing the non-refractive physical reference, along the normal vector of the adhesive groove. (Reference gradient) This is the grayscale gradient of the pre-captured image of the adhesive tray area without any transparent lampshade obstructing the view. The reference gradient was acquired under conditions identical to the current image acquisition conditions, including an illumination intensity of 800 lux, a ring light source angle of 45 degrees, a camera exposure time of 10 milliseconds, and a gain of 1.0. Current gradient During the assembly process The grayscale gradient of the image containing the transparent lampshade was acquired. The image gradient was calculated using a 3×3 Sobel operator, calculating the gradient components in the x and y directions separately, and then synthesizing the gradient magnitude.

[0021] S303 uses the normal offset as the independent variable to perform a cross-correlation matching operation along the normal direction between the current gradient and the reference gradient, generating a matching response sequence that reflects the degree of local texture deformation. For each possible phase offset... In the normal scan area The product of the current gradient and the reference gradient is numerically integrated using the trapezoidal rule to obtain the matched response value corresponding to the phase shift. The matched response values ​​corresponding to all phase shifts constitute a matched response sequence. (Phase shift) The range of values ​​is The search step size is 0.1 pixels.

[0022] S304 selects the phase offset that causes the matched response sequence to reach its peak value. The phase offsets corresponding to each circumferential parameter position are combined to form a refractive phase field characterizing the degree of refractive distortion of the transparent component. The formula for calculating the refractive phase field is: ; in Circumferential parameters The refracted phase field value at the location, in pixels; It is the phase offset, measured in pixels; This means taking the option that maximizes the integral result. The number of sampling points for the circumferential parameter is 500, and the sampling interval is 0.002, corresponding to the uniform sampling of the normalized circumferential parameter.

[0023] S401 sequentially splices together the pose deviation (characterizing rigid body displacement), the edge deformation (characterizing component elasticity), and the rubber strip deformation (characterizing the compression state of the seal) to form candidate particle states. The calculation formula for the particle states is: ; in It is the first The state vector of each particle; It is a 6-dimensional pose deviation vector, containing 3 translation components and 3 rotation components. The translation component is in meters and the rotation component is in radians. It is an 8-dimensional edge deformation vector, with units of meters; It is a 100-dimensional gel bar variable vector, with units in meters.

[0024] S402 spatially maps the edge deformation based on a preset deformation basis function and superimposes it onto the outer edge baseline of the lampshade to obtain a local contour line characterizing the bending shape under stress. The deformation basis function adopts a fourth-order Fourier basis function, in the form of: .

[0025] The process of spatial mapping is to calculate Circumferential parameters are obtained The edge deformation value at the specified location is measured in meters. This edge deformation value is then superimposed onto the reference line along the normal direction of the lampshade's outer edge reference line to obtain the local contour line. .

[0026] The outer edge baseline of the lampshade in the circumferential parameters Unit normal vector at the location The calculation method is exactly the same as that for the glue groove normal vector. First, calculate the three-dimensional tangential vector of the outer edge baseline of the lampshade. Then, by cross-product of the camera optical axis unit vector After normalization, the calculation formula is as follows: .

[0027] S403 performs Lie algebraic exponential mapping on the pose deviation to generate a transformation matrix representing the spatial pose change. Using this transformation matrix, it performs a coordinate system composite transformation on the initial pose and the local contour line, outputting the outer edge deformation line. The formula for calculating the outer edge deformation line is: ; in It is the first The outer edge deformation line corresponding to each particle is a three-dimensional coordinate sequence in the robot's base coordinate system; It is the initial pose transformation matrix of the lampshade, which is a 4×4 homogeneous transformation matrix; It is the pose deviation vector The result of the Lie algebraic exponential mapping is a 4×4 homogeneous transformation matrix; It is a 6-dimensional Lie algebra vector The operation of converting to a 4×4 antisymmetric matrix, if ,but The Lie algebraic exponential mapping is calculated using the Rodriguez formula. For the rotation part, if the rotation vector is... Its module length is Then the rotation matrix ,in It is a 3×3 identity matrix. The translation part is directly... .

[0028] When the rotation vector magnitude When, rotation matrix ,in It is a 3×3 identity matrix, and the translation part is directly... No complex matrix operations are required.

[0029] S501 extracts the distribution difference between the refracted phase field and the predicted phase field generated by the current particle prediction, uses the fluctuation variance to perform feature normalization on the distribution difference, and performs logarithmic integration to obtain the visual cost used to suppress visual misalignment errors. Predicted phase field The simulation calculation using ray tracing is as follows: 1. For the circumferential parameters... Center line point of the glue tank Along the normal vector of the glue groove Direction is taken as normal offset , obtain the surface points of the glue tank 2. From the camera's optical center towards Emit light rays and calculate the deformation line of the light rays relative to the outer edge of the lampshade. intersection 3. Based on the refractive index of the lampshade and thickness 4. Calculate the direction of light refraction inside the lampshade; 5. Calculate the position of the refracted light reaching the surface of the glue tank. The predicted phase offset is obtained. 5. Traverse all normal offsets (unit: pixels); and circumferential parameters The predicted phase field is obtained. Volatility and variance Circumferential parameters The historical variance of the refracted phase field is calculated from the refracted phase field data of the most recent 100 production batches. Circumferential weights. Circumferential parameters The weighting coefficients are set to 2.0 at the four corners of the lampshade and 1.0 at the other positions.

[0030] lampshade thickness Extracted from the product's 3D CAD model as circumferential parameters Normal thickness of the outer edge of the lampshade, in meters. Predicted phase offset. The formula for converting meters to pixels is: ,in The average focal length of the camera, in pixels; For the surface of the glue tank The z-axis coordinate in the camera coordinate system, in meters.

[0031] S502 performs quadratic energy calculations on edge deformation based on the stiffness matrix and introduces visual weights to derive the deformation cost used to constrain excessive bending deformation of the lampshade. Stiffness matrix This is an 8×8 matrix representing the elastic stiffness characteristics of the lampshade edge, obtained through finite element analysis. The parameters for the finite element analysis are: lampshade material is polycarbonate, elastic modulus is 2.2 gigapascals, Poisson's ratio is 0.37, mesh size is 1 mm, and the boundary constraint is that the lampshade is fixed at its six mounting points. Visual weights. The value is The reciprocal of Joule. The formula for calculating deformation cost is: .

[0032] S503 extracts the volume difference between the current cross-sectional area of ​​the adhesive strip and the containment area under the current compression, performs a weighted square integral on the volume difference, and obtains the volume cost used to prevent localized adhesive overflow. Current adhesive strip cross-sectional area Circumferential parameters The cross-sectional area of ​​the rubber strip before compression, in square meters, is obtained from parameters provided by the rubber strip supplier. Current compression level. It is the first circumferential parameters in the state of a single particle The compression of the rubber strip at the point, in meters, is measured through the outer edge deformation line. Centerline of the glue tank The three-dimensional normal distance between them is calculated. The cross-section of the glue tank is U-shaped, and the accommodating area is... The calculation formula is: ,in This refers to the width of the glue tank, in meters. The radius of the fillet at the bottom of the glue tank, in meters, is obtained from the CAD model. Geometric weights. The value is The reciprocal of the fourth power of meters. The formula for calculating volume cost is: .

[0033] Rubber strip compression For the outer edge deformation line points To the center line of the glue tank The three-dimensional normal distance is calculated using the following formula: ,in This represents the dot product operation of three-dimensional vectors.

[0034] S504 performs a joint penalty fusion of visual cost, deformation cost, and volumetric cost to obtain the fitness function. The formula for calculating the fitness function is: ; in It is the first The fitness function value of a particle; the smaller the value, the better the state of the particle.

[0035] S601 calculates the state differences of all particle states from the population mean state, and extracts the population dispersion, a metric representing the convergence of the particle distribution, based on Mahalanobis distance. The formula for calculating the population dispersion is: ; in It is the first Group dispersion of a particle swarm; It is the total number of particles in the particle swarm, with a value of 50; It is the first The generation The state vector of each particle; It is the first The mean state vector of the particle swarm is calculated using the following formula: ; It is the first The covariance matrix of the particle swarm state vector is calculated using the following formula: ; This represents the square of the Mahalanobis distance. When the covariance matrix... When it is strange, add Multiply by the identity matrix for regularization.

[0036] S602 uses group dispersion as a moderating variable to calculate the inertia weight and cognitive factor used to maintain the exploratory distribution, and calculates the social factor used to guide group convergence. The formulas for calculating the inertia weight, cognitive factor, and social factor are as follows: ; in It is the first The inertial weight of the generation; It is the first Cognitive factors of the generation; It is the first Social factors of the era. The initial parameters of the particle swarm are set as follows: initial velocity range is... The initial particle state distribution range is the pose deviation translation component. Meter, rotational component Curvature, edge deformation Meter, rubber strip deformation rice.

[0037] S603 utilizes inertia weights, cognitive factors, social factors, and environmental random variables, combining individual optimality and global optimality, to synchronously and iteratively update the velocity vector and particle state of the current individual, thereby avoiding local minima. The update formulas for the velocity vector and particle state are as follows: ; ; in It is the first The generation The velocity vector of each particle; It is the first The generation The velocity vector of each particle; and The range of values ​​is Uniformly distributed random numbers; It is the first The generation The individual optimal state vector of a particle, that is, the state in which the fitness function value of the particle is the smallest in history. It is the first The global optimal state vector of the particle swarm is the state with the minimum fitness function value throughout the entire history of the swarm. Upper limits are set on particle velocities, with the translation component having an upper limit of [value missing]. The upper limit of the rotation component per meter iteration is For each iteration of the radian, the upper limit of edge deformation is... For each iteration, the upper limit of the strip-shaped variable is... For each iteration, when the speed exceeds the upper limit, the corresponding upper limit value is taken.

[0038] When the updated particle state exceeds the preset range, a boundary truncation operation is performed, and the truncation range for the pose deviation translation component is [value missing]. Meters, the rotational component cutoff range is The radius and edge deformation cutoff range are as follows Meters, the cutoff range of the rubber strip shape is rice.

[0039] After multiple sampling cycles of iterative updates, S604 outputs the globally optimal value that minimizes the penalty value of the fitness function, and this value is selected as the optimal particle. The iteration terminates when the number of iterations reaches 100, or when the change in the globally optimal fitness function value over 20 consecutive iterations is less than [a certain value]. When any termination condition is met, the iteration stops, and the current globally optimal state vector is output. As the optimal particle.

[0040] S701 calculates the Jacobian matrix of the predicted phase field corresponding to the optimal particle and the pose deviation, establishing a sensitivity mapping relationship between visual features and spatial pose changes. (Jacobi matrix) It is a 500×6 matrix, where 500 represents the number of sampling points for the circumferential parameters, and 6 represents the dimension of the pose deviation vector. The Jacobian matrix is ​​calculated using numerical differentiation. Each component Add tiny increments The increment of the translation component is meters, the increment of the rotational component is radians, the predicted phase field after calculating the increment ,in For the first The unit basis vectors. The Jacobian matrix of the nth unit basis vector. Line number The formula for calculating column elements is: ; in It is the first Normalized circumferential parameter values ​​for each sampling point.

[0041] S702 introduces a damping matrix characterizing motion stagnation and performs a regularized inverse operation on the Jacobian matrix. Combining this with the phase difference between the predicted and refracted phase fields, the pose correction is generated using a damped least squares descent method. The formula for calculating the pose correction is: ; in It is the pose correction vector, which is a 6-dimensional Lie algebra vector; It is a 500×500 phase weight diagonal matrix, where the diagonal elements are the circumferential weights at each sampling point. ; It is a 6×6 damped diagonal matrix, and the diagonal elements corresponding to the translational degrees of freedom take values ​​of 100,0 ... The diagonal elements corresponding to the rotational degrees of freedom take values ​​of .

[0042] The S703 performs a manifold space exponential mapping operation on the pose correction amount and composites it onto the current pose to obtain the target pose used to correct spatial misalignment. The formula for calculating the target pose is: ; in It is the corrected target pose transformation matrix; It is the current pose transformation matrix; It is the pose correction vector The Lie algebra exponent mapping results are calculated using the same method as in S403.

[0043] The S704 combines pose correction with control step size to convert it into end-effector velocity. Joint velocities are then derived through pseudo-inverse kinematics calculations of the robotic arm to smoothly adjust the robot's assembly system's spatial posture. The formula for calculating joint velocity is: ; in It is the joint velocity vector of the robotic arm, measured in radians per second; It is the current joint angle of the robotic arm. The kinematic pseudo-inverse matrix under; This controls the step size, with a value of 0.05 seconds. It is the velocity vector at the end of the robotic arm, with the translation component in meters per second and the rotation component in radians per second.

[0044] pseudo-inverse matrix of robotic arm kinematics The Moore-Penrose pseudo-inverse matrix is ​​calculated using the following formula: ,in For the robotic arm at joint angles The following is a 6×n Jacobian matrix, where n is the number of robotic arm joints; The damping coefficient has a value of [value missing]. ; It is an n×n identity matrix.

[0045] S801, by introducing edge deformation as a physical offset constraint, searches the optimization space for the transformation matrix that minimizes the joint penalty value of the closure cost and volume cost representing the spatial contact gap. This matrix is ​​then used as the final pose to eliminate assembly tolerance uncertainties. The formula for calculating the final pose is: ; in It is the final pose transformation matrix; It is the pose transformation matrix to be optimized; It is the vector of edge deformation contained in the optimal particle; In the transformation matrix and edge deformation The outer edge deformation line below; This represents the square of the weighted Euclidean distance. It is a 500×500 distance-weighted diagonal matrix, with diagonal elements corresponding to the circumferential weights. Consistent; In the transformation matrix and edge deformation Below, circumferential parameters The compression amount of the adhesive strip at that location. The pose range of the final pose optimization space is the translation component. Meter, rotational component Radius. The integral of the joint penalty value is calculated using the trapezoidal rule.

[0046] The S802 extracts the attitude difference vector from the current pose to the final pose, and performs incremental interpolation by combining variables representing the pressing progress with the kinematic pseudo-inverse matrix to generate a continuous pressing path that avoids impact interference. The calculation formula for the continuous pressing path is: ; in The pressing progress is The joint angle vector of the robotic arm at that time; It is the current joint angle vector of the robotic arm; It is the homogeneous transformation matrix The steps for converting to a 6-dimensional Lie algebra vector are as follows: 1. Extract the rotation part of the transformation matrix. Translation section 2. Calculate the rotation vector of the rotation matrix. 3. Calculate the Lie algebra components corresponding to the translation part using the inverse operation of the Rodriguez formula; ,in 4. Concatenate to obtain a 6-dimensional Lie algebra vector. . It is a pressing progress variable, and its value range is... ,when When corresponding to the current pose, when The time corresponds to the final pose. The update step size for the pressing progress is 0.01, and the interpolation period is 10 milliseconds.

[0047] When the rotation angle corresponding to the rotation matrix At that time, the Lie algebra components corresponding to the translation part ,in is the translation vector of the transformation matrix.

[0048] The S803 drive robot assembly system advances along a continuous pressing path until full scale is reached, simultaneously saving the refracted phase field, optimal particle and terminal pose at the end of assembly, forming a visual record for process status backtracking. Full scale refers to the pressing progress. At this point, the lampshade is fully pressed into the glue groove. The auxiliary verification condition for proper assembly is that the pressure sensor at the end of the robotic arm detects a value of 50 Newtons. When the pressure reaches the threshold and the pressing progress reaches full scale, the assembly is considered complete. Visual recordings are stored in JSON format, and the data structure includes an assembly timestamp, a 500-dimensional array of refractive phase fields, a 114-dimensional optimal particle state vector, and a 4×4 terminal pose transformation matrix.

[0049] The pressure sensor is installed between the end flange of the robotic arm and the lamp holder. The sampling frequency is 100 Hz, the filter window size is 5, and the moving average filter is used to eliminate high-frequency noise.

[0050] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for controlling the assembly of automotive parts based on visual positioning feedback, applied to a robotic assembly system including a camera, for pressing a polycarbonate transparent lampshade into a peripheral adhesive groove of a polypropylene shell, characterized in that... include: Images of the periphery of the polypropylene shell and the polycarbonate transparent lampshade are obtained, and a circumferential coordinate system is established and bound to the robot assembly system to unify the normal reference of the lampshade and the glue groove. In the circumferential coordinate system, the refracted phase field generated when the polycarbonate transparent lampshade approaches the peripheral adhesive groove is extracted to remove the refraction artifact interference caused by the transparent curved surface; Construct a particle state that characterizes assembly bias and elastic deformation. The particle state includes pose deviation, edge deformation amount and rubber strip shape change. The particle state is used to generate an outer edge deformation line that characterizes the actual contact contour. Based on the refracted phase field, a fitness function is established by combining the deformation cost characterizing physical stress and the volume cost characterizing adhesive overflow. Perform adaptive particle swarm visual feedback search and output the optimal particle that minimizes the fitness function; The optimal direction is calculated based on the optimal particle, the pose correction amount is generated, and the robot assembly system is controlled to adjust the spatial posture. The final pose is obtained based on the closure cost and volume cost representing the assembly gap. The robot assembly system is then controlled to complete the positioning along the continuous pressing path, and a visual record is formed simultaneously.

2. The automotive component assembly control method based on visual positioning feedback according to claim 1, characterized in that, Establishing a circumferential coordinate system bound to the robot assembly system includes: Obtain the camera intrinsic parameters of the camera, and the camera extrinsic parameters of the robot assembly system to the camera; Extract the center line of the housing glue groove and the outer edge reference line of the lampshade from the acquired images to characterize the assembly reference; By using the camera's extrinsic parameters and intrinsic parameters to construct a perspective mapping relationship, the target coordinates in the space where the robot assembly system is located are projected onto the camera's two-dimensional plane to eliminate the hand-eye coordinate difference; The tangential variation trend of the center line of the glue groove in the shell along the circumferential parameters is extracted, and after orthogonal rotation and mold length normalization, a glue groove normal vector is generated to indicate the direction of force on the glue groove. Using the circumferential parameter as a unified variable, the center line of the housing groove and the outer edge reference line of the lampshade are aligned to the same circumferential coordinate system.

3. The automotive component assembly control method based on visual positioning feedback according to claim 2, characterized in that, Extracting the refracted phase field includes: The normal scanning area used to eliminate background environmental interference is defined according to the inner and outer wall boundaries of the surrounding adhesive tank; Within the normal scanning area, along the normal vector of the glue groove, the current gradient representing the current visual features and the reference gradient representing the non-refractive physical reference are extracted respectively. Using the normal offset as the independent variable, a cross-correlation matching operation along the normal is performed on the current gradient and the reference gradient to generate a matching response sequence that reflects the degree of local texture deformation. A phase offset that causes the matching response sequence to reach its peak value is selected, and the phase offsets corresponding to each of the circumferential parameter positions are combined to form the refractive phase field used to characterize the degree of refractive distortion of the transparent part.

4. The automotive component assembly control method based on visual positioning feedback according to claim 3, characterized in that, Constructing a particle state and generating an outer edge deformation line using the particle state includes: The pose deviation used to characterize rigid body displacement, the edge deformation amount used to characterize component elasticity, and the rubber strip deformation amount used to characterize the compression state of seal are sequentially spliced ​​together to form the candidate particle state. The edge deformation amount is spatially mapped based on the preset deformation basis function and superimposed on the outer edge baseline of the lampshade to obtain a local contour line characterizing the bending shape under stress. The pose deviation is subjected to Lie algebraic exponential mapping to generate a transformation matrix characterizing the spatial pose change. The initial pose and the local contour line are then subjected to coordinate system composite transformation using the transformation matrix to output the outer edge deformation line.

5. The automotive component assembly control method based on visual positioning feedback according to claim 4, characterized in that, Establish a fitness function, including: Extract the distribution difference between the refracted phase field and the predicted phase field generated by the current particle prediction, use the fluctuation variance to perform feature normalization on the distribution difference and perform logarithmic integration to obtain the visual cost used to suppress visual misalignment error. Based on the stiffness matrix, a quadratic energy calculation is performed on the edge deformation amount, and a visual weight is introduced to obtain the deformation cost used to constrain the excessive bending deformation of the lampshade. Extract the volume difference between the current cross-sectional area of ​​the adhesive strip and the accommodating area under the current compression amount, and perform a weighted square integral on the volume difference to obtain the volume cost used to prevent local adhesive overflow; The fitness function is obtained by jointly penalizing and fusing the visual cost, the deformation cost, and the volume cost.

6. The automotive component assembly control method based on visual positioning feedback according to claim 5, characterized in that, Perform adaptive particle swarm visual feedback search, including: Calculate the state difference between all the particle states and the population mean state, and extract the population dispersion based on the Mahalanobis distance metric to characterize the degree of convergence of the particle distribution; Using the group dispersion as a moderating variable, calculate the inertia weight and cognitive factor used to maintain the exploration distribution, and calculate the social factor used to guide the group convergence. By utilizing the inertial weight, the cognitive factor, the social factor, and environmental random variables, and combining individual optimality with global optimality, the velocity vector of the current individual and the particle state are updated synchronously and iteratively to avoid local minima. After iterative updates over multiple sampling periods, the output that minimizes the penalty value of the fitness function is the global optimum, which is then used as the optimal particle.

7. The automotive component assembly control method based on visual positioning feedback according to claim 6, characterized in that, Generating pose correction values ​​and controlling the adjustment of spatial orientation, including: Calculate the Jacobian matrix of the predicted phase field corresponding to the optimal particle with respect to the pose deviation, and establish the sensitivity mapping relationship of visual features with spatial pose changes. A damping matrix characterizing motion stagnation is introduced to perform a regularized inverse operation on the Jacobian matrix. The pose correction is generated by combining the phase difference between the predicted phase field and the refracted phase field using the damped least squares descent method. The pose correction amount is subjected to manifold space exponential mapping operation and then composited onto the current pose to obtain the target pose for correcting spatial misalignment. The pose correction amount is combined with the control step size and converted into an end-effector velocity. The joint velocity is obtained through pseudo-inverse kinematics calculation of the robotic arm to control the robot assembly system to smoothly adjust its spatial posture.

8. The automotive component assembly control method based on visual positioning feedback according to claim 7, characterized in that, Obtain the final pose and complete the positioning along the continuous pressing path, including: Under the condition of introducing the edge deformation amount as a physical bias constraint, the transformation matrix that minimizes the joint penalty value of the closure cost and the volume cost that characterizes the spatial contact gap is found in the optimization space, and it is used as the final pose to eliminate assembly tolerance uncertainty. Extract the attitude difference vector from the current pose to the final pose, combine it with the variable representing the pressing progress and the kinematic pseudo-inverse matrix to perform incremental interpolation operation, and generate the continuous pressing path that avoids impact interference. The robot assembly system is driven to advance along the continuous pressing path until the progress reaches full scale, and the refracted phase field, the optimal particle and the terminal pose at the end of the assembly are saved simultaneously to form the visual record for process status backtracking.

Images