Hub plug-in tightening system and method based on 3D vision and digital twinning

Abstract

The invention discloses a hub plug-in tightening system and method based on 3D vision and digital twinning, and the system comprises an image collection module which is used for executing a semi-global matching algorithm and depth calculation; the feature extraction module is used for extracting local geometric features and global features of the spatial points through a feature extraction model based on deep learning; the pose matching and compensating module is used for calculating an overall error, a rotation error, an angle error and a translation error between an actual hole pose and a theoretical mounting pose; the digital twinning mapping module is used for constructing a digital twinning model and correcting the compensation control quantity based on virtual and real errors; and the torque tightening control module is used for generating a motor output torque based on the motor rotation error and correcting the motor output torque. According to the system, the system can adapt to passenger vehicle hubs of different specifications, the plug-in holes and other circular characteristics of the hubs are effectively distinguished, meanwhile, error accumulation can be effectively restrained, and the running stability of the system is improved.

Description

Technical Field

[0001] This invention relates to the field of wheel assembly, and more specifically to a wheel hub plug tightening system and method based on 3D vision and digital twin. Background Technology

[0002] The wheel hub insertion tightening system based on 3D vision and digital twin is an industrial precision assembly equipment that integrates 3D vision positioning technology, deep learning feature extraction, digital twin virtual-real mapping and adaptive torque control. The wheel hub insertion tightening system is mainly composed of an image acquisition module, a feature extraction module, a pose matching and compensation module, a digital twin mapping module and a torque tightening control module. The wheel hub insertion tightening system is widely used in the wheel hub insertion assembly process in the automotive manufacturing industry.

[0003] In practical applications, existing technologies are susceptible to damage from oil stains and edge wear, leading to large deviations in the estimation of parameters such as hole center and normal vector, and making it difficult to adapt to various wheel hub specifications. Typically, position compensation is an open-loop control, which makes it difficult to correct real-time fluctuations such as wheel hub deformation and mechanical clearance, resulting in incorrect position compensation direction and amplified error accumulation. Furthermore, it is prone to plug jamming, over-tightening, under-tightening, or even misjudging the position of plug holes, thereby affecting the tightening accuracy and assembly quality of wheel hub plugs. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, the present invention aims to provide a wheel hub insert tightening system and method based on 3D vision and digital twins, in order to solve the problems of insert jamming, over-tightening, under-tightening, and even misjudgment of insert hole position in the existing systems. The system of the present invention improves the adaptability to wheel hubs of different specifications while ensuring accuracy, and meets the requirements of actual use scenarios.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The present invention relates to a wheel hub insert tightening system based on 3D vision and digital twin, the system comprising:

[0007] Module 1: Image acquisition module, used to execute semi-global matching algorithm and depth calculation, generate three-dimensional spatial point calculation results and form a spatial point cloud set, and at the same time calculate the coordinates of the wheel hub center;

[0008] Module 2: Feature Extraction Module, used to normalize the coordinates of spatial points, extract local geometric features and global features of spatial points through a deep learning-based feature extraction model, and filter the set of plug-in hole parameters through a classification function;

[0009] Module 3: Pose Matching and Compensation Module, used to calculate the overall error, rotation error, angle error and translation error between the actual hole pose and the theoretical installation pose, and generate the compensation control quantity corresponding to each plug-in hole;

[0010] Module 4: Digital Twin Mapping Module, used to establish a real-time state mapping between the physical system and the virtual system of the wheel hub insert tightening; construct a digital twin model, and correct the compensation control quantity based on the virtual-real error;

[0011] Module 5: Torque Tightening Control Module, used to establish a servo motor dynamics model, convert the corrected compensation control quantity into a target rotation reference angle for the motor; generate the motor output torque based on the motor rotation error, and correct the motor output torque.

[0012] The present invention provides a wheel hub insert tightening method based on 3D vision and digital twin, the method comprising:

[0013] Step 1: Acquire left and right eye images using two cameras respectively, and calculate pixel disparity using a semi-global matching algorithm; obtain the coordinates in the camera coordinate system using a 3D spatial point calculation function; calculate the wheel hub center coordinates using a wheel hub center calculation function;

[0014] Step 2: The new spatial point coordinates are scaled using a normalization function to obtain normalized spatial point coordinates; local geometric features and global features of the spatial points are extracted using a feature extraction model; and the plug-in hole parameter set is screened out using a classification function based on the global features.

[0015] Step 3: Calculate the overall error, rotation error, angle error, and translation error between the actual hole pose and the theoretical installation pose; based on the angle error, normal vector, and translation error, obtain the compensation control amount corresponding to each plug-in hole through the pose compensation function;

[0016] Step 4: Generate the predicted state of the virtual system using a digital twin model based on the physical system's state vector and the compensation control quantity; then, correct the compensation control quantity using a correction function based on the virtual-real error.

[0017] Step 5: Convert the corrected compensation control quantity into the target rotation reference angle of the motor using a conversion function; obtain the motor output torque based on the angle error and angular velocity error generated by the motor rotation and combined with the adaptive control function; optimize the motor output torque using an optimization function to obtain the corrected motor output torque.

[0018] Furthermore, the function for calculating the hub center in step one is as follows:

[0019]

[0020] in, The coordinates of the wheel hub center; A collection of spatial point clouds; For the first The coordinates of a point in space;

[0021] The coordinates of a point in space are centered to obtain new coordinates. The centering function is shown below:

[0022]

[0023] in, These are the new coordinates of the spatial point.

[0024] Furthermore, the nearest neighbor set in step two is as follows:

[0025]

[0026] in, For the first A set of nearest neighbors of a normalized spatial point; for Nearest neighbor algorithm function, The nearest neighbor number; These are the normalized coordinates of a point in space.

[0027] The multi-scale feature aggregation function is shown below:

[0028]

[0029] in, For the first The spatial point at the th Local geometric features at various scales; For MLP mapping functions, These are the learnable weight parameters of the MLP mapping function; The coordinates of the points in the space after normalization For the center of the ball, A three-dimensional spherical neighborhood with radius . The first in the nearest neighbor set The coordinates of a normalized point in space.

[0030] Furthermore, the classification function in step two is as follows:

[0031]

[0032] in, For classification functions; This is a set of parameters for the plug-in hole, including the hole center. Hole radius normal vector , This represents the total number of insertion holes; Index for plug-in holes.

[0033] Furthermore, the pose compensation function in step three is as follows:

[0034]

[0035] in, For the first Compensation control amount for each plug-in hole; This is the gain coefficient for rotational error compensation; This is the gain coefficient for translation error compensation; For the first Angle error of each plug hole; It is the normal vector; This represents the translation error.

[0036] Furthermore, the digital twin model in step four is as follows:

[0037]

[0038] in, The state of the digital twin system at the next moment; For digital twin state transition functions; Let the physical system state vector be... It is a time variable; Compensation control quantity.

[0039] Furthermore, the optimization function in step five is as follows:

[0040]

[0041] in, This is the corrected motor output torque; For proportional gain; This is the differential gain; For integral gain; This refers to the angular error caused by the rotation of the motor. This refers to the angular velocity error caused by the rotation of the motor. This is the error feedback gain matrix; For virtual and real errors, It is a time variable.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] The system of this invention proposes a feature extraction module that enables the system to adapt to different specifications of passenger car wheel hubs without the need to train a separate model for each model, greatly improving the system's versatility and reducing the debugging cost of multi-variety mixed production lines. At the same time, the feature extraction model can comprehensively capture the local details, overall outline and global relative position of the plug-in hole, effectively distinguishing the plug-in hole from other circular features of the wheel hub.

[0044] The system of this invention proposes a digital twin mapping module, which enables the system to predict and compensate through real-time virtual-real mapping and state prediction functions. It can predict changes in the state of the physical system, thereby reducing response lag errors under dynamic operating conditions. At the same time, based on the correction of compensation control quantity of virtual-real error, a closed-loop optimization is formed, which further reduces the initial compensation error, effectively suppresses error accumulation, and improves the stability of system operation. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the system's workflow in this invention;

[0047] Figure 2 This is a schematic diagram of the image acquisition module architecture of the system in this invention;

[0048] Figure 3 This is a schematic diagram of the feature extraction module architecture of the system in this invention;

[0049] Figure 4 This is a schematic diagram of the pose matching and compensation module architecture of the system in this invention;

[0050] Figure 5 This is a schematic diagram of the digital twin mapping module architecture of the system in this invention;

[0051] Figure 6 This is a schematic diagram of the torque tightening control module architecture in the system of this invention. Detailed Implementation

[0052] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0053] Reference Figure 1 As shown, the present invention provides a wheel hub insert tightening system based on 3D vision and digital twins, the system comprising:

[0054] Module 1: Image acquisition module, used to execute semi-global matching algorithm and depth calculation, generate three-dimensional spatial point calculation results and form a spatial point cloud set, and at the same time calculate the coordinates of the wheel hub center;

[0055] First, left and right eye images were acquired using two cameras, respectively. Pixel disparity was then calculated using a semi-global matching algorithm. The cameras used were 640×480 industrial cameras with a specific focal length. Baseline distance Optical center coordinates in camera intrinsic parameters , Number of plug-in holes The hub of the wheel, the total number of spatial points in the point cloud set. For example, the semi-global matching algorithm is shown below:

[0056]

[0057]

[0058]

[0059] in, For pixel parallax, All are pixel coordinates in the image plane; The x-coordinate of the left eye image, for example, the x-coordinate of the acquired left eye image. ; The x-coordinate of the right eye image, for example, the x-coordinate of the acquired right eye image. ; Image for the left eye; Image for the right eye;

[0060] Substituting the data into the formula, we obtain the pixel disparity:

[0061]

[0062] Based on pixel disparity, a depth calculation function is constructed to calculate the spatial depth value. The depth calculation function is shown below:

[0063]

[0064] in, This represents the spatial depth value. The focal length of the camera; Baseline distance; For pixel parallax;

[0065] Substituting the data into the formula, we obtain the spatial depth value:

[0066]

[0067] Secondly, based on the spatial depth value, a 3D spatial point calculation function is constructed to obtain the coordinates in the camera coordinate system. The 3D spatial point calculation function is shown below:

[0068]

[0069]

[0070] in, For image coordinates The x-coordinate of the back projection in the camera coordinate system; For image coordinates The ordinate of the back projection onto the camera coordinate system; and These are the coordinates of the optical center in the camera's intrinsic parameters, representing the coordinates of the intersection of the camera's optical axis and the image plane in the image coordinate system. The focal length of the camera;

[0071] Substituting the data into the formula, we obtain the x-coordinate and y-coordinate in the camera coordinate system:

[0072]

[0073] Based on the coordinates in the camera coordinate system, the coordinates of the spatial points are obtained as follows:

[0074]

[0075] in, For the first The coordinates of a point in space; for A spatial point in Coordinates on the axis; for A spatial point in Coordinates on the axis; for A spatial point in Coordinates on the axis;

[0076] Based on the x-coordinate and y-coordinate in the camera coordinate system, the coordinates of the spatial point are obtained:

[0077]

[0078] The coordinates of all spatial points are summarized to generate a spatial point cloud set, as shown below:

[0079]

[0080] in, A collection of spatial point clouds; The total number of spatial points in the point cloud set;

[0081] Finally, a function to calculate the hub center coordinates is constructed, as shown below:

[0082]

[0083]

[0084] in, The coordinates of the wheel hub center;

[0085] Substituting the data into the formula, we obtain the coordinates of the wheel hub center:

[0086]

[0087] The coordinates of a point in space are centered to obtain new coordinates. The centering function is shown below:

[0088]

[0089] in, For the new spatial coordinates;

[0090] Substituting the data into the formula, we obtain the new spatial point coordinates:

[0091]

[0092] The camera in the image acquisition module is a Basler acA640-120gm, and the camera is installed at a height of 300~800mm from the surface of the wheel hub, with the field of view completely covering the wheel hub and no mechanical structure obstructing it.

[0093] Module 2: Feature Extraction Module, used to normalize the coordinates of spatial points, extract local geometric features and global features of spatial points through a deep learning-based feature extraction model, and filter the set of plug-in hole parameters through a classification function;

[0094] First, the new spatial point coordinates are scale-normalized to obtain normalized spatial point coordinates, enabling the model to adapt to different sizes of passenger car wheel hubs without needing to train a separate model for each type of wheel hub. The normalization function is shown below:

[0095]

[0096] in, These are the normalized coordinates of a point in space. For the new spatial coordinates; The new spatial point coordinates are defined by the Euclidean norm.

[0097] Substituting the new spatial point coordinates into the formula, we obtain the normalized spatial point coordinates:

[0098]

[0099] Secondly, construct a feature extraction model based on deep learning:

[0100] Define a deep learning feature mapping function, as shown below:

[0101]

[0102] in, For mapping functions, These are network weight parameters; For feature dimensions; This represents the total number of 3D points contained in the point cloud set.

[0103] For each normalized spatial point coordinate, construct Nearest neighbor sets, serving as the basis for local feature extraction, are shown below:

[0104]

[0105] in, For the first A set of nearest neighbors of a normalized spatial point; for Nearest neighbor algorithm function, To determine the nearest neighbor number, based on industry experience in point cloud processing, a parameter search range is set. The wheel hub point cloud dataset is then divided into a training set. The parameter combinations within the search range are traversed to train the feature extraction model. Finally, the nearest neighbor number with the best overall performance on the validation set is selected, using the plug hole classification accuracy and hole center prediction error as evaluation metrics. These are the normalized coordinates of a point in space.

[0106] Extracting local geometric features based on nearest neighbor sets can preserve key features within the neighborhood and avoid feature blurring caused by average aggregation.

[0107]

[0108] in, For the first Local geometric features of a normalized spatial point; For MLP mapping functions, These are the learnable weight parameters of the MLP mapping function; For the neighboring region The normalized coordinates of a point in space;

[0109] Based on local geometric features, multi-scale feature aggregation is performed to avoid the limitations of single-scale features, improve the recognizability of plug hole features, and reduce interference from other circular features of the wheel hub. The multi-scale feature aggregation function is shown below:

[0110]

[0111] in, For the first The spatial point at the th Local geometric features at various scales; For MLP mapping functions, For the learnable weight parameters of the MLP mapping function, and For different sets of network parameters; The coordinates of the points in the space after normalization For the center of the ball, For a three-dimensional spherical neighborhood with radius, the Euclidean norm distribution of spatial points is statistically analyzed using the centered and normalized hub point cloud data to determine the basic range of the radius. Then, different combinations of radii are tested using the controlled variable method.

[0112] The aggregated multi-scale features can avoid information loss caused by feature superposition and ensure that features at different scales can all play a role.

[0113]

[0114] in, For the first Multi-scale features of spatial points after multi-scale feature aggregation; For multi-scale feature fusion operators;

[0115] The multi-scale features of each spatial point are summarized to obtain the global features:

[0116]

[0117] in, Global features of spatial points; The total number of spatial points in the point cloud set;

[0118] For example, constructing the coordinates of each normalized spatial point. Nearest neighbor set, using spherical neighborhoods of 3 scales. Multi-scale feature aggregation and summarization are performed to obtain global features:

[0119]

[0120] Finally, a classification function is constructed to filter out the set of plug-in hole parameters based on global features. The classification function is shown below:

[0121]

[0122] in, For classification functions; This is a set of parameters for the insertion hole, including the hole center. Hole radius normal vector , This represents the total number of insertion holes; Index for plug-in holes;

[0123] Substitute the data into the formula to filter out the set of plug-in hole parameters. The data is shown in the table below:

[0124]

[0125] Construct the loss function, as shown below:

[0126]

[0127]

[0128]

[0129]

[0130] in, This is the total loss function; For classification loss; For the first The actual classification labels of each spatial point; For the model to predict the first The probability that a spatial point is a plug-in hole; For the center regression loss of the hole; For the first The actual center coordinates of each plug-in hole were obtained by manual annotation. The loss is the normal vector regression loss; For the first The true normal vector of each plug hole, a unit vector obtained from the CAD model; The weighting coefficients for the regression loss at the hole center; The weighting coefficients for the normal vector regression loss are defined. First, an orthogonal experimental table of weight combinations is established. Then, using the total loss value and the actual insertion hole positioning error as indicators, the optimal combination is selected, and thus the result is determined. and ;

[0131] The network structure of the feature extraction model in deep learning mainly consists of an input layer (normalized spatial point coordinates), a sampling grouping layer (neighborhood construction), a local feature extraction layer, a multi-scale aggregation layer, a global feature layer, and a classification and regression output layer (classification function).

[0132] The activation function of the feature extraction model in deep learning is the ReLU function in the sampling grouping layer and the multi-scale aggregation layer, and the Sigmoid function is used in the classification and regression output layer.

[0133] The edge computing unit in the feature extraction module uses NVIDIA Jetson AGX Xavier, the high-speed storage unit uses Samsung 870EVO 1TB SSD, the industrial-grade motherboard unit uses AIMB-785G4, and the data transmission interface card uses Intel I225-LM Gigabit Ethernet.

[0134] Module 3: Pose Matching and Compensation Module, used to calculate the overall error, rotation error, angle error and translation error between the actual hole pose and the theoretical installation pose, and generate the compensation control quantity corresponding to each plug-in hole;

[0135] First, the theoretical installation position of the preset plugin:

[0136]

[0137] in, Install the pose homogeneous transformation matrix for the plug-in theory; This is the theoretical rotation matrix; The coordinates of the actual hole center;

[0138] With theoretical rotation matrix For example, the homogeneous transformation matrix of the plug-in's theoretical installation pose is obtained:

[0139]

[0140] Based on the plug-in hole parameter set, construct the actual hole pose matrix:

[0141]

[0142]

[0143] in, For the first The actual pose homogeneous transformation matrix of each plug hole; for The actual rotation matrix of each insertion hole contains any unit vector orthogonal to the normal vector. The unit vector obtained by the cross product of any unit vector and the normal vector. normal vector Hole center ;

[0144] by Taking the plug-in hole as an example, substituting the data, we obtain the actual pose homogeneous transformation matrix:

[0145]

[0146] Secondly, an overall error calculation function is constructed to calculate the overall error between the actual hole pose and the theoretical installation pose, which is used to comprehensively evaluate the degree of pose deviation of the plug-in hole. The overall error calculation function is as follows:

[0147]

[0148] in, For the first The homogeneous transformation matrix of the pose error of each plug hole represents the overall error between the actual hole pose and the theoretical installation pose.

[0149] Yi Yi Taking the plug-in hole as an example, substituting the data, we obtain the homogeneous transformation matrix of pose error:

[0150]

[0151] Construct a rotation error calculation function to calculate the rotation error. The rotation error calculation function is shown below:

[0152]

[0153] in, For the first The rotation error matrix of each insertion hole represents the rotation error;

[0154] by Taking the plug-in hole as an example, substituting the data, we obtain the rotation error matrix:

[0155]

[0156] Based on the rotation error matrix, an angle error calculation function is constructed, which can quantify the rotation deviation into an angle value that can be directly used for compensation. The angle error calculation function is shown below:

[0157]

[0158] in, For the first Angle error of each plug hole; The trace of the rotation error matrix;

[0159] by Taking the insertion hole as an example, substituting the rotation error matrix data into the formula, we obtain the angle error:

[0160]

[0161] Construct a translation error calculation function to calculate the translation error. The translation error calculation function is shown below:

[0162]

[0163] in, This is the translation error; The coordinates of the actual hole center;

[0164] by Taking the insertion hole as an example, substituting the data into the formula, we obtain the translation error:

[0165]

[0166] Finally, based on the angle error, normal vector, and translation error, a pose compensation function is constructed, as shown below:

[0167]

[0168] in, For the first The compensation control amount of each plug-in hole is used to drive the actuator to compensate for the positional error and achieve precise plug-in installation. For example, rotation error compensation gain coefficient. ; For example, the translation error compensation gain coefficient. First, a reasonable range is determined using control theory, then fine-tuning is performed on the actual production line to determine the final range. and ;

[0169] Substitute the data into the formula to calculate the compensation control value:

[0170]

[0171] The motion controller unit in the pose matching and compensation module is a Delta Tau PMAC2-PC / 104, the servo driver unit is a Panasonic MINAS A6 series, and the encoder interface unit is a Beckhoff EL5001.

[0172] Module 4: Digital Twin Mapping Module, used to establish a real-time state mapping between the physical system and the virtual system of the wheel hub insert tightening; construct a digital twin model, and correct the compensation control quantity based on the virtual-real error;

[0173] First, a real-time state mapping function is constructed between the physical system and the virtual system for tightening the wheel hub plug. This facilitates time-series prediction of the digital twin model and ensures that the model can track the state changes of the physical system in real time. The real-time state mapping function is shown below:

[0174]

[0175] in, Let be the state vector of the physical system, representing time . The real-time state of the physical entity; It is a time variable; Center of the hole; This is for angular error; Compensation control quantity;

[0176] Will Substitute the data of the plug hole into the formula and take the time. ,get Physical system state vector of the plug-in hole:

[0177]

[0178] Secondly, a digital twin model is constructed. Based on the physical system's state vector and compensation control variables, a virtual system predictive state is generated to achieve state prediction of the wheel hub tightening system, enabling advance control of the tightening process. The digital twin model is shown below:

[0179]

[0180] in, The state of the digital twin system at the next moment; For digital twin state transition functions;

[0181] Substituting the data into the digital twin model, we obtain the state of the digital twin system at the next moment:

[0182]

[0183] Based on the state of the digital twin system and the state vector of the physical system at the next moment, the virtual-real error is calculated, which reflects the prediction accuracy of the digital twin model. The virtual-real error calculation function is shown below:

[0184]

[0185] in, This is the error between real and virtual values;

[0186] Substituting the physical system state vector and the digital twin system state data at the next moment into the formula, we obtain the virtual-real error:

[0187]

[0188] Finally, based on the virtual and real errors, a correction function is constructed to correct the compensation control quantity, thereby achieving closed-loop correction of the compensation control quantity and improving the accuracy of the compensation control. The correction function is shown below:

[0189]

[0190]

[0191] in, This is the corrected compensation control value; For the first Compensation control amount for each plug-in hole; This is the error feedback gain matrix, which includes the hole center in the corresponding physical state. Feedback gain coefficient For example, the feedback gain coefficient at the center of the aperture. Angular error in the corresponding physical state Feedback gain coefficient For example, the feedback gain coefficient of angle error. The initial compensation control quantity in the corresponding physical state Feedback gain coefficient For example, the feedback gain coefficient of the initial compensation control quantity. The system identifies the virtual and real error transmission model between the physical and virtual systems, determines the basic range of the feedback gain coefficient, and then uses the convergence speed of the virtual and real errors and the accuracy of the corrected compensation control quantity as indicators to conduct iterative trial screwing on the production line, gradually adjust the feedback gain coefficient, and finally determine the feedback gain coefficient.

[0192] Substituting the data into the formula, we obtain the corrected compensation control value:

[0193]

[0194] The network structure of the digital twin model consists of a mechanism layer computation layer (virtual system prediction state) and a data-driven correction layer. The data-driven correction layer includes a time-series input layer, a GRU time-series layer, an MLP mapping layer, and a correction output layer.

[0195] The activation function of the digital twin model is the Tanh function in the GRU time series layer and the ReLU function in the MLP mapping layer;

[0196] Activation function of digital twin model:

[0197]

[0198] in, The total loss function for digital twins; The center of the hole in the physical system state; The virtual hole center was predicted by the digital twin model; This refers to the angular error under the physical system conditions. The virtual angle error is predicted by the digital twin model. This refers to the corrected compensation control quantity under the physical system state; The virtual corrected compensation control quantity is predicted by the digital twin model; This is the weighting coefficient for the hole center position error; This is the angle error weighting coefficient; Error weighting coefficient for compensation control quantity;

[0199] The digital twin mapping module uses an IPC-610L industrial computer unit, an NVIDIA A10 timing computing accelerator card unit, a Beckhoff EL6692 real-time communication unit, and a SiemensTecnomatix Plant Simulation 2201 digital twin software platform.

[0200] Module 5: Torque Tightening Control Module, used to establish a servo motor dynamics model, convert the corrected compensation control quantity into a target rotation reference angle for the motor; generate the motor output torque based on the motor rotation error, and correct the motor output torque;

[0201] First, construct the servo motor dynamic model, as shown below:

[0202]

[0203] in, This is the equivalent rotational inertia of the motor and load. It is the viscous damping coefficient; For example, load torque during the tightening process of the plug. ; This refers to the output torque of the motor. This refers to the angular acceleration of the motor. This refers to the angular velocity of the motor. The angular displacement of the motor output shaft is used; the equivalent rotational inertia of the motor and load is employed. Viscous damping coefficient servo motors;

[0204] Secondly, a conversion function is constructed to convert the corrected compensation control quantity into the target rotation reference angle of the motor. The conversion function is shown below:

[0205]

[0206] in, The reference angle for the target rotation of the motor; This is the corrected compensation control value; This is the normal vector of the insertion hole;

[0207] by Taking the data of the plug-in hole as an example, substitute the data into the formula to obtain the target rotation reference angle of the motor:

[0208]

[0209] Based on the target rotation reference angle of the motor, the angular error and angular velocity error generated by the motor rotation are calculated using the following functions:

[0210]

[0211]

[0212] in, This refers to the angular error caused by the rotation of the motor. This represents the actual rotation angle of the motor. This refers to the angular velocity error caused by the rotation of the motor. The reference angular velocity for the target rotation of the motor; This is the actual angular velocity of the motor;

[0213] The actual rotation angle of the servo motor is collected by an industrial encoder. The actual angular velocity of the servo motor And preset the target rotation reference angular velocity of the servo motor. Substitute the data into the formula to calculate the angular error and angular velocity error generated by the rotation of the servo motor:

[0214]

[0215]

[0216] Based on the angular and angular velocity errors generated by the motor rotation, an adaptive control function is constructed to obtain the motor output torque. The adaptive control function is shown below:

[0217]

[0218] in, This refers to the output torque of the motor. For example, proportional gain. ; For example, differential gain. ; For example, integral gain. The ZN tuning method is adopted, and the basic value of the gain is calculated through the step response curve of the motor. Then, the gain coefficient is finally determined by tightening and adjusting the hub plug.

[0219] Substitute the data into the formula to calculate the motor output torque:

[0220]

[0221] This means adding to the base tightening torque. The torque;

[0222] Finally, based on the virtual and real errors, the motor output torque is optimized to obtain the corrected motor output torque, achieving the final precise optimization of the wheel hub plug tightening torque. This ensures that both the positional accuracy and torque accuracy of the wheel hub plug tightening meet engineering requirements. The optimization function is shown below:

[0223]

[0224]

[0225] in, This is the corrected motor output torque;

[0226] Substituting the data into the formula, the corrected motor output torque is calculated:

[0227]

[0228] This means adding to the base tightening torque. The torque;

[0229] Among them, the servo motor unit in the torque tightening control module adopts the Panasonic MINAS A6 series, the torque sensor unit adopts the HBM T40B, the motor reducer unit adopts the Harmonic Drive CSF-17-50-2A-GR, and the torque amplifier unit adopts the HBM MGCplus.

[0230] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A wheel hub insert tightening system based on 3D vision and digital twin, characterized in that, The system includes: Module 1: Image acquisition module, used to execute semi-global matching algorithm and depth calculation, generate three-dimensional spatial point calculation results and form a spatial point cloud set, and at the same time calculate the coordinates of the wheel hub center; Module 2: Feature Extraction Module, used to normalize the coordinates of spatial points, extract local geometric features and global features of spatial points through a deep learning-based feature extraction model, and filter the set of plug-in hole parameters through a classification function; Module 3: Pose Matching and Compensation Module, used to calculate the overall error, rotation error, angle error and translation error between the actual hole pose and the theoretical installation pose, and generate the compensation control quantity corresponding to each plug-in hole; Module 4: Digital Twin Mapping Module, used to establish a real-time state mapping between the physical system and the virtual system of the wheel hub insert tightening; construct a digital twin model, and correct the compensation control quantity based on the virtual-real error; Module 5: Torque Tightening Control Module, used to establish a servo motor dynamics model, convert the corrected compensation control quantity into a target rotation reference angle for the motor; generate the motor output torque based on the motor rotation error, and correct the motor output torque.

2. A method for tightening wheel hub inserts based on 3D vision and digital twins, characterized in that, Includes the following steps: Step 1: Acquire left and right eye images using two cameras respectively, and calculate pixel disparity using a semi-global matching algorithm; The coordinates in the camera coordinate system are obtained through a 3D spatial point calculation function; the coordinates of the wheel hub center are calculated through a wheel hub center calculation function. Step 2: The new spatial point coordinates are scaled using a normalization function to obtain normalized spatial point coordinates; local geometric features and global features of the spatial points are extracted using a feature extraction model; and the plug-in hole parameter set is screened out using a classification function based on the global features. Step 3: Calculate the overall error, rotation error, angle error, and translation error between the actual hole pose and the theoretical installation pose; based on the angle error, normal vector, and translation error, obtain the compensation control amount corresponding to each plug-in hole through the pose compensation function; Step 4: Generate the predicted state of the virtual system using a digital twin model and based on the physical system's state vector and compensation control variables; The compensation control quantity is corrected by using a correction function and based on the virtual and real errors; Step 5: Convert the corrected compensation control quantity into the target rotation reference angle of the motor using a conversion function; based on the angle error and angular velocity error generated by the motor rotation and combined with the adaptive control function, obtain the motor output torque; By optimizing the function, the motor output torque is optimized to obtain the corrected motor output torque.

3. The method according to claim 2, characterized in that, The function for calculating the hub center in step one is as follows: in, The coordinates of the wheel hub center; A collection of spatial point clouds; For the first The coordinates of a point in space; The coordinates of a point in space are centered to obtain new coordinates. The centering function is shown below: in, These are the new coordinates of the spatial point.

4. The method according to claim 2, characterized in that, The nearest neighbor set in step two is shown below: in, For the first A set of nearest neighbors of a normalized spatial point; for Nearest neighbor algorithm function, The nearest neighbor number; These are the normalized coordinates of a point in space. The multi-scale feature aggregation function is shown below: in, For the first The spatial point at the th Local geometric features at various scales; For MLP mapping functions, These are the learnable weight parameters of the MLP mapping function; The coordinates of the points in the space after normalization For the center of the ball, A three-dimensional spherical neighborhood with radius . The first in the nearest neighbor set The coordinates of a normalized point in space.

5. The method according to claim 2, characterized in that, The classification function in step two is as follows: in, For classification functions; This is a set of parameters for the plug-in hole, including the hole center. Hole radius normal vector , This represents the total number of insertion holes; Index for plug-in holes.

6. The method according to claim 2, characterized in that, The pose compensation function in step three is as follows: in, For the first Compensation control amount for each plug-in hole; This is the gain coefficient for rotational error compensation; This is the gain coefficient for translation error compensation; For the first Angle error of each plug hole; It is the normal vector; This represents the translation error.

7. The method according to claim 2, characterized in that, The digital twin model in step four is shown below: in, The state of the digital twin system at the next moment; For digital twin state transition functions; Let the physical system state vector be... It is a time variable; Compensation control quantity.

8. The method according to claim 2, characterized in that, The optimization function in step five is as follows: in, This is the corrected motor output torque; For proportional gain; This is the differential gain; For integral gain; This refers to the angular error caused by the rotation of the motor. This refers to the angular velocity error caused by the rotation of the motor. This is the error feedback gain matrix; For virtual and real errors, It is a time variable.

Images