A Deep Learning-Based Control Method and System for Wheel Hub Bolt Press-fitting

Abstract

This invention discloses a deep learning-based control method and system for wheel hub bolt press-fitting, belonging to the field of automotive wheel hub assembly technology. The method first pre-trains a deep learning model for rotating target detection based on Gaussian Wasserstein distance. After wheel hub positioning, image acquisition, bolt hole pose calculation, and coaxial alignment, a staged closed-loop press-fitting strategy combining admittance compliant control and active disturbance rejection control is adopted. This invention improves the accuracy and robustness of bolt hole pose detection, mitigates the rigid impact of press-fitting, suppresses multi-source nonlinear disturbances in the press-fitting process, ensures consistent press-fitting quality, adapts to the needs of multi-model mixed-line production, and can be widely applied to automated press-fitting scenarios for automotive wheel hub bolts.

Description

Technical Field

[0001] This invention relates to the field of automotive wheel hub assembly technology, specifically to a wheel hub bolt press-fitting control method and system based on deep learning. Background Technology

[0002] In the production of modern automotive chassis and wheel hub assembly, wheel hub bolt pressing is a key process that directly affects vehicle driving safety. The pressing quality directly affects the overall vehicle driving stability, and also determines the production line's operating cycle and product yield. As the automotive manufacturing industry develops towards automation and flexibility, the requirements for positioning accuracy and pressing stability in wheel hub bolt pressing technology continue to increase, while also needing to adapt to the rapid switching needs of multi-model mixed-line production.

[0003] Currently, wheel hub bolt pressing is conventionally completed using mechanical positioning fixtures in conjunction with presses and servo actuators. Although some solutions introduce industrial vision technology to assist in obtaining bolt hole position information, the overall process still relies on rigid positioning, fixed trajectory execution, and a single control strategy to complete the pressing process, which has core technical defects in practical applications.

[0004] On the one hand, existing visual positioning solutions lack the accuracy and robustness for detecting near-circular wheel hub bolt holes. Conventional rotating target detection and visual positioning methods are prone to problems such as discontinuous angle regression and pose jumps. In complex industrial environments such as metal surface reflection and background interference, it is difficult to output stable and high-precision bolt hole pose information, and cannot provide reliable pose guidance for press-fitting operations. At the same time, they rely heavily on customized rigid positioning fixtures, making it difficult to adapt to the mixed-line production needs of multiple wheel hub models.

[0005] On the other hand, the existing press-fitting control schemes lack compliance and anti-disturbance capabilities. Conventional single control strategies are difficult to adapt to the multi-stage dynamic characteristics of the press-fitting process. During the contact stage, the system is prone to generating impact loads due to excessive rigidity, which can cause problems such as thread scratches and orifice damage. At the same time, it cannot effectively suppress nonlinear disturbances such as friction and load changes during the press-fitting process, making it difficult to ensure the consistency of press-fitting quality of different batches of workpieces.

[0006] Therefore, it is necessary to design a deep learning-based control method and system for wheel hub bolt press-fitting. Summary of the Invention

[0007] The purpose of this invention is to provide a wheel hub bolt press-fitting control method and system based on deep learning, so as to solve the problems of insufficient visual positioning accuracy and robustness, weak production line flexibility and adaptability, poor press-fitting control compliance and anti-disturbance ability, and difficulty in ensuring press-fitting quality consistency in the existing wheel hub bolt press-fitting technology mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Firstly, a deep learning-based method for controlling the press-fitting of wheel hub bolts is provided, comprising the following steps:

[0010] S1 Pre-trained Rotating Target Detection Deep Learning Model: Construct a rotating target detection deep learning model for wheel hub bolt hole detection. Convert the bolt hole annotations in the training samples into a two-dimensional Gaussian distribution. Calculate the regression loss of the rotating bounding box using the Wasserstein distance between the two two-dimensional Gaussian distributions to complete model training.

[0011] S2 Wheel Hub Positioning and Image Acquisition: The wheel hub to be pressed is transported and positioned to the pressing station through the station positioning unit, and the workpiece image including the wheel hub bolt holes is acquired through the vision acquisition unit;

[0012] S3 calculates the spatial pose of the wheel hub bolt holes: The acquired workpiece image is input into the model pre-trained by S1, and the rotated bounding box of the bolt hole is output; Sub-pixel level edge extraction and ellipse fitting are performed on the rotated bounding box region, and combined with the geometric prior of the circumferentially distributed wheel hub bolt holes, the topological geometric rigid body consistency of the fitting result is checked, and the spatial pose of the bolt hole is calculated.

[0013] S4 Actuator Coaxial Alignment: Based on the affine transformation matrix between the image pixel coordinate system determined in advance through eye-to-hand calibration and the physical coordinate system of the press-fitting actuator, the spatial pose of the target bolt hole obtained in S3 is converted into position and angle compensation amounts, driving the press head of the press-fitting actuator to be coaxially aligned with the target bolt hole.

[0014] S5 phased press-fit closed-loop control: Start the press-fit process and collect the press-fit force and displacement of the press-fit actuator and the real-time visual pose signal of the vision acquisition unit in real time;

[0015] When the press contact force is detected to exceed the preset contact threshold, the system switches to admittance compliance control mode, which treats the press head as a virtual mass-damping-stiffness second-order system and generates a position correction amount based on the real-time force deviation.

[0016] When the bolt is detected to have entered the preset guide depth, the system switches to the linear active disturbance rejection control mode, and the position correction amount is superimposed on the preset press-fit displacement trajectory to generate the desired tracking trajectory and complete the tracking.

[0017] When the pressing displacement reaches the target depth and the pressing force reaches the rated load, switch to constant force steady-state control mode, maintain the preset constant force until the pressure holding is completed, and end the bolt pressing.

[0018] As a further technical solution of the present invention, S1 specifically includes the following:

[0019] S1.1 Model Architecture Construction: A single-stage anchor-free rotating target detection deep learning model is constructed, consisting of a backbone network, a feature fusion layer, and a decoupled detection head. The backbone network adopts the CSPDarknet structure to extract multi-scale features from the input image. The feature fusion layer adopts the FPN+PAMet structure to achieve efficient fusion of deep semantic features and shallow detail features of the image. The decoupled detection head is divided into independent classification and regression branches. The regression branch outputs five core parameters of the bolt hole rotation bounding box: the pixel coordinates of the bounding box center, the width and height of the bounding box, and the rotation angle of the bounding box.

[0020] S1.2 Construction of Training Sample Set: Collect image samples covering bolt holes of different wheel hub models, label the core parameters of the corresponding rotated bounding box for each bolt hole in the sample, and generate a labeled dataset; perform data augmentation processing such as image normalization, random flipping, and brightness perturbation on the labeled dataset to expand the sample richness, and finally construct the model training sample set;

[0021] S1.3 Two-dimensional Gaussian mapping of rotated bounding boxes: The rotated bounding boxes predicted by the model and the real rotated bounding boxes labeled in the training samples are mapped to two-dimensional Gaussian distributions respectively; where the center coordinates of the rotated bounding box correspond to the mean vector of the two-dimensional Gaussian distribution, and the width, height and rotation angle of the bounding box correspond to the covariance matrix of the two-dimensional Gaussian distribution, thus converting the rigid rotated bounding box into a continuous Gaussian probability distribution.

[0022] S1.4 Regression loss calculation based on Wasserstein distance: Based on the Wasserstein distance between two two-dimensional Gaussian distributions, a regression loss function for rotating bounding boxes is constructed; first, the Wasserstein distance between the Gaussian distribution corresponding to the predicted result and the Gaussian distribution corresponding to the true label is calculated, and then a smooth and continuous loss value is generated through nonlinear mapping.

[0023] S1.5 Iterative Model Training: Cross-entropy loss is used as the classification loss for the model's classification branch, and the loss constructed based on Wasserstein distance is used as the regression loss for the regression branch. These are combined to generate the model's total loss function. A stochastic gradient descent optimizer is used to iteratively update all the model's weight parameters through the backpropagation algorithm until the total loss function converges, thus completing the model's pre-training.

[0024] As a further technical solution of the present invention, S2 specifically includes the following:

[0025] S2.1 Hub Conveying and Station Positioning: The multi-station linear conveyor of the station positioning unit conveys the hub to be pressed to the servo rotary table of the pressing station; the photoelectric sensor detects the hub's arrival signal and triggers the conveyor to stop; then the servo rotary table is controlled to perform initial circumferential positioning of the hub, completing the fixing and locking of the hub at the pressing station;

[0026] S2.2 Image Acquisition Triggering and Lighting Control: After receiving the wheel hub positioning signal, the industrial controller synchronously triggers the surround industrial camera group of the vision acquisition unit and the surface light source lighting system; it controls the surface light source lighting system to output a large area of ​​shadowless diffused light to uniformly illuminate the surface of the wheel hub.

[0027] S2.3 Synchronous Acquisition of Workpiece Images: Control the surround industrial camera group to synchronously acquire color workpiece images containing all bolt holes of the wheel hub from multiple perspectives, and transmit the acquired image data to the industrial controller in real time;

[0028] S2.4 Workpiece Image Preprocessing: The industrial controller performs preprocessing operations on the acquired color workpiece images in sequence to generate images that meet the model input requirements. Specifically, this includes converting the color image into a single-channel grayscale image, using median filtering to denoise the grayscale image to suppress salt-and-pepper noise, enhancing the edge gradient information of the bolt holes through nonlinear grayscale transformation, and finally extracting the region of interest containing all bolt holes based on the outer contour features of the wheel hub.

[0029] As a further technical solution of the present invention, S3 specifically includes the following:

[0030] S3.1 Model Inference Output Rotation Bounding Box: Input the pre-processed workpiece image into the pre-trained deep learning model for rotating target detection, and output the core parameters of the rotation bounding box corresponding to each bolt hole in the image through model inference;

[0031] S3.2 Subpixel-level edge extraction: For each image region corresponding to a rotated bounding box, firstly, candidate edge pixels are extracted by gradient calculation, and then the position of the candidate edge pixels is optimized by subpixel-level interpolation through quadratic curve fitting to obtain more accurate edge coordinates;

[0032] S3.3 Ellipse Fitting and Hole Center Calculation: For the sub-pixel edge point set corresponding to each bolt hole, the least squares method is used to fit the ellipse model, and the coefficients of the ellipse model are obtained. Then, based on the fitted ellipse model, the center coordinates of the bolt hole are calculated.

[0033] S3.4 Topological Geometry Rigid Body Consistency Verification: First, obtain the number of bolt holes on the wheel hub to be press-fitted. Based on the geometric prior of the circumferentially uniformly distributed bolt holes, calculate the wheel hub geometric rotation center corresponding to the center coordinates of all bolt holes in sequence. Then, calculate the circumferential angle of each bolt hole relative to the wheel hub geometric rotation center based on the ellipse fitting result. If the circumferential angle deviation of a single bolt hole exceeds the preset threshold, it is judged as a false detection result and discarded. The wheel hub geometric rotation center and corresponding pose parameters are recalculated based on the fitting results of the remaining bolt holes.

[0034] S3.5 Bolt Hole Spatial Pose Output: The final output is the spatial pose parameters of each bolt hole, including the pixel coordinates of the bolt hole center, the circumferential angle relative to the geometric rotation center of the hub, and the hole diameter.

[0035] As a further technical solution of the present invention, S4 specifically includes the following:

[0036] S4.1 Hand-to-Hand Calibration Parameter Call: Calls the calibration parameters that have been determined in advance through the Eye-to-Hand calibration method, including the affine transformation relationship between the image pixel coordinate system and the physical coordinate system of the press-fitting actuator, and the circumferential angle mapping relationship between the image pixel coordinate system and the physical coordinate system of the servo rotary table;

[0037] S4.2 Pose Conversion and Compensation Calculation: Based on the affine transformation relationship obtained from calibration, the center pixel coordinates of the target bolt hole are converted into three-dimensional position coordinates in the physical coordinate system of the press-fitting actuator, and the position compensation of the press head is calculated; based on the circumferential angle mapping relationship obtained from calibration, the circumferential angle of the target bolt hole is converted into the rotation angle compensation of the servo rotary table.

[0038] S4.3 Coaxial Alignment Drive: Based on the calculated position compensation amount, the press head of the press actuator is driven to move directly above the target bolt hole; based on the calculated rotation angle compensation amount, the servo rotary table is driven to rotate and adjust, so that the central axis of the press head is completely coincident with the central axis of the target bolt hole, thus completing the coaxial alignment.

[0039] As a further technical solution of the present invention, S5 specifically includes the following:

[0040] S5.1 Pressing Process Start-up and Real-time Signal Acquisition: The industrial controller starts the pressing process and controls the pressing head of the pressing actuator to move downward at a preset feed speed; the pressing force signal and displacement signal are acquired in real time at a preset sampling frequency through the pressure sensor and displacement sensor integrated at the end of the pressing head, and the visual pose signal of the pressing head and bolt hole is acquired in real time through the vision acquisition unit.

[0041] S5.2 Contact Stage Admittance Compliance Control: When the real-time pressing force exceeds the preset contact threshold, the system switches to admittance compliance control mode; the pressing head of the pressing actuator is equivalent to a second-order compliant system including virtual inertia, virtual damping, and virtual stiffness. A system dynamic model is established, and the deviation between the real-time collected pressing force and the preset guiding contact force is used as input. The real-time position correction of the pressing head is obtained through the dynamic model, and the feed position of the pressing head is dynamically adjusted until the bolt is detected to enter the preset guiding depth, thus ending this stage of control.

[0042] S5.3 Linear Active Disturbance Rejection Control during Pressing Stage: When the bolt is detected to have entered the preset guide depth, the system switches to linear active disturbance rejection control mode. First, the position correction amount of the admittance compliance control output is superimposed on the preset press-fit displacement trajectory to generate the final desired tracking trajectory. Then, a dynamic model of the press-fit actuator is established, and the nonlinear friction, dead zone, parameter fluctuations and external resistance disturbances inside the system are uniformly attributed to the total disturbance of the system. A third-order linear extended state observer is constructed to track and estimate the displacement state, velocity state and total disturbance of the system in real time. Based on the tracking error between the desired tracking trajectory and the actual state, combined with the real-time estimate of the total disturbance, the control quantity of the servo drive is calculated to control the feed action of the press head in real time and perform disturbance compensation until the press-fit displacement reaches the target depth and the press-fit force reaches the rated load, thus ending this stage of control.

[0043] S5.4 Constant Force Steady-State Control during the Holding Pressure Stage: When the pressing displacement reaches the target depth and the pressing force reaches the rated load, the system switches to constant force steady-state control mode. Based on the real-time collected deviation between the pressing force and the rated load, the position fine-tuning amount of the pressing head is calculated, and the position of the pressing head is dynamically adjusted. At the same time, through the linear expansion state observer, the equivalent disturbance caused by the elastic recovery of the material and the release of residual stress is estimated and compensated in real time to maintain the pressing force stable at the rated load. Until the holding pressure time reaches the preset value, the constant force steady-state control ends, and the pressing head is controlled to return at the preset speed to complete the entire bolt pressing process.

[0044] As a further technical solution of the present invention, the conversion of the rotated bounding box into a two-dimensional Gaussian distribution is specifically as follows:

[0045] For five sets of parameters of an arbitrary rotated bounding box Mapped to a two-dimensional Gaussian distribution ,in:

[0046] Mean vector:

[0047] Covariance matrix:

[0048]

[0049] In the formula: This is the x-coordinate of the bounding box center in the image pixel coordinate system; The ordinate of the bounding box center in the image pixel coordinate system; This represents the width of the bounding box; This represents the height of the bounding box; The rotation angle of the bounding box, with a value range of 1. .

[0050] As a further technical solution of the present invention, the regression loss calculation of the rotated bounding box is specifically as follows:

[0051] Calculate the predicted Gaussian distribution Compared with the true Gaussian distribution The square of the two-dimensional Wasserstein distance between them :

[0052]

[0053] In the formula: For L2 norm operations; This is the trace operation of a matrix; The rotated bounding box predicted by the model; The actual rotated bounding boxes labeled in the training samples;

[0054] Calculate the regression loss of the rotated bounding box :

[0055]

[0056]

[0057] In the formula: For smoothing hyperparameters.

[0058] As a further technical solution of the present invention, when a pressing force signal is detected... Exceeding the preset contact threshold When switching to admittance compliance control mode, perform the following operations:

[0059] Establish the second-order dynamic equations for the virtual mass-damping-stiffness system of the pressure head:

[0060]

[0061] In the formula: For virtual inertia parameters, For virtual damping parameters, For virtual stiffness parameters, This is the position correction amount output by the admittance model. for The first derivative, for The second derivative, Preset guiding contact force.

[0062] Based on the above dynamic equations, according to the real-time force deviation The position correction amount at the current sampling time is calculated. The output is sent to the press-fitting actuator to correct the feed position of the press head in real time;

[0063] Until a displacement signal is detected Once the preset guidance depth threshold is reached, the admittance compliance control mode ends.

[0064] Secondly, a deep learning-based wheel hub bolt press-fitting control system is provided, including a reinforced four-column reaction force bearing frame, a station positioning mechanism, a vision acquisition mechanism, a press-fitting execution mechanism, a feeding mechanism, and an industrial controller.

[0065] The reinforced four-column reaction load-bearing frame includes four vertically parallel precision guide columns, a top beam, and a bottom base. The top beam is fixedly connected to the top of the four precision guide columns, and the bottom base is fixedly connected to the bottom of the four precision guide columns, forming a closed-loop load-bearing body.

[0066] The workstation positioning mechanism includes a multi-station linear conveyor and a servo rotary table. The multi-station linear conveyor is fixedly installed in the middle area of ​​the reinforced four-column reaction bearing frame. The servo rotary table is fixedly installed at the pressing station of the multi-station linear conveyor, and the workpiece bearing surface of the servo rotary table is set upward.

[0067] The vision acquisition mechanism includes multiple industrial cameras and multiple sets of LED array panel lights. The multiple industrial cameras are symmetrically fixedly installed on the inner sidewalls of four precision guide columns, forming a camera group surrounding the pressing station. The lenses of all industrial cameras face the workpiece bearing surface of the servo rotary table. The multiple sets of LED array panel lights are fixedly installed on the lower surface of the top crossbeam, and the light-emitting surfaces of all LED array panel lights face the workpiece bearing surface of the servo rotary table.

[0068] The press-fitting actuator includes a dual servo hydraulic power source and a press head. The dual servo hydraulic power source is fixedly installed at the center of the top crossbeam. The power output end of the dual servo hydraulic power source is fixedly connected to the press head through a precision guide rod. The press head is located above the servo rotary table. The lower end of the press head integrates a pressure sensor and a displacement sensor.

[0069] The feeding mechanism is an industrial robot, which is installed on the feeding side of the reinforced four-column reaction bearing frame; the industrial controller is electrically connected to the multi-station linear conveyor, servo rotary table, industrial camera, LED array panel light, dual servo hydraulic power source, industrial robot, pressure sensor, and displacement sensor respectively.

[0070] The industrial controller is configured to execute the deep learning-based wheel hub bolt press-fitting control method described in the first aspect.

[0071] Compared with existing technologies, the beneficial effects of this deep learning-based wheel hub bolt press-fitting control method and system are:

[0072] By constructing a deep learning model for rotating target detection based on Gaussian distribution and Wasserstein distance, the problems of angle jumps and regression discontinuities that are prone to occur in the detection of near-circular bolt holes are solved from the bottom layer. Combined with a surround-type multi-view vision acquisition architecture and a topological geometric rigid body consistency verification mechanism, it can effectively eliminate the influence of metal surface reflection, shadow occlusion, and background interference on the detection results in complex industrial environments, and output stable and continuous bolt hole pose information, providing high-precision and high-robust pose guidance for subsequent pressing operations, and avoiding the defects of pose abrupt changes and insufficient detection robustness that are prone to occur in traditional vision positioning methods.

[0073] The closed-loop press-fitting control architecture adopts a phased adaptive switching approach. During the press-fitting contact stage, the actuator is given compliant adaptation capability through admittance compliant control, which can effectively alleviate the rigid impact at the moment of press-fitting contact and avoid problems such as scratches, stripping, and hole damage between bolts and workpiece holes. During the pressing stage, the active disturbance rejection control realizes real-time observation and compensation of the nonlinear characteristics of the system and complex external disturbances. Combined with the constant force steady-state control during the holding pressure stage, it can effectively suppress the impact of various uncertain factors on the press-fitting quality throughout the entire press-fitting process, ensure the consistency and stability of the press-fitting process of different batches of workpieces, and reduce the risk of equipment overload and abnormal shutdown.

[0074] A fully automated closed-loop control method and system was constructed, encompassing wheel hub positioning, image acquisition, pose calculation, coaxial alignment, and staged press-fitting. This system enables precise press-fitting of wheel hub bolts without relying on high-precision rigid positioning fixtures. It can quickly adapt to wheel hub workpieces of different models and with different hole distributions, significantly improving the flexibility of the production line and meeting the switching needs of multi-model mixed-line production. At the same time, the fully automated closed-loop control reduces manual intervention, effectively improving the production line's cycle time and yield rate, and reducing the production line's operation and maintenance costs. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0076] Figure 2 This is a schematic diagram of the deep learning model network structure for rotating target detection according to the present invention;

[0077] Figure 3 This is a schematic diagram of the phased press-fit closed-loop control principle of the present invention;

[0078] Figure 4 This is a schematic diagram of the system physical architecture of the present invention;

[0079] In the picture: 1. LED array panel light; 2. Dual servo hydraulic system; 3. Industrial robot; 4. Precision guide column; 5. Camera group; 6. Servo rotary table. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] Please see the appendix Figure 1 -Appendix Figure 3 The present invention provides an embodiment of a deep learning-based wheel hub bolt press-fit control method, comprising the following steps:

[0082] S1 Pre-trained Rotating Target Detection Deep Learning Model: Construct a rotating target detection deep learning model for wheel hub bolt hole detection. Convert the bolt hole annotations in the training samples into a two-dimensional Gaussian distribution. Calculate the regression loss of the rotating bounding box using the Wasserstein distance between the two two-dimensional Gaussian distributions to complete model training.

[0083] S2 Wheel Hub Positioning and Image Acquisition: The wheel hub to be pressed is transported and positioned to the pressing station through the station positioning unit, and the workpiece image including the wheel hub bolt holes is acquired through the vision acquisition unit;

[0084] S3 calculates the spatial pose of the wheel hub bolt holes: The acquired workpiece image is input into the model pre-trained by S1, and the rotated bounding box of the bolt hole is output; Sub-pixel level edge extraction and ellipse fitting are performed on the rotated bounding box region, and combined with the geometric prior of the circumferentially distributed wheel hub bolt holes, the topological geometric rigid body consistency of the fitting result is checked, and the spatial pose of the bolt hole is calculated.

[0085] S4 Actuator Coaxial Alignment: Based on the affine transformation matrix between the image pixel coordinate system determined in advance through eye-to-hand calibration and the physical coordinate system of the press-fitting actuator, the spatial pose of the target bolt hole obtained in S3 is converted into position and angle compensation amounts, driving the press head of the press-fitting actuator to be coaxially aligned with the target bolt hole.

[0086] S5 phased press-fit closed-loop control: Start the press-fit process and collect the press-fit force and displacement of the press-fit actuator and the real-time visual pose signal of the vision acquisition unit in real time;

[0087] When the press contact force is detected to exceed the preset contact threshold, the system switches to admittance compliance control mode, which treats the press head as a virtual mass-damping-stiffness second-order system and generates a position correction amount based on the real-time force deviation.

[0088] When the bolt is detected to have entered the preset guide depth, the system switches to the linear active disturbance rejection control mode, adds the position correction amount to the preset press-fit displacement trajectory, generates the desired tracking trajectory, and completes the tracking.

[0089] When the pressing displacement reaches the target depth and the pressing force reaches the rated load, switch to constant force steady-state control mode, maintain the preset constant force until the pressure holding is completed, and end the bolt pressing.

[0090] Furthermore, in another embodiment, S1 specifically includes the following:

[0091] S1.1 Model Architecture Construction: A single-stage anchor-free rotating target detection deep learning model is constructed, consisting of a backbone network, a feature fusion layer, and a decoupled detection head. The backbone network adopts the CSPDarknet structure to extract multi-scale features from the input image. The feature fusion layer adopts the FPN+PAMet structure to achieve efficient fusion of deep semantic features and shallow detail features of the image. The decoupled detection head is divided into independent classification and regression branches. The regression branch outputs five core parameters of the bolt hole rotation bounding box: the pixel coordinates of the bounding box center, the width and height of the bounding box, and the rotation angle of the bounding box.

[0092] S1.2 Construction of Training Sample Set: Collect image samples covering bolt holes of different wheel hub models, label the core parameters of the corresponding rotated bounding box for each bolt hole in the sample, and generate a labeled dataset; perform data augmentation processing such as image normalization, random flipping, and brightness perturbation on the labeled dataset to expand the sample richness, and finally construct the model training sample set;

[0093] S1.3 Two-dimensional Gaussian mapping of rotated bounding boxes: The rotated bounding boxes predicted by the model and the real rotated bounding boxes labeled in the training samples are mapped to two-dimensional Gaussian distributions respectively; where the center coordinates of the rotated bounding box correspond to the mean vector of the two-dimensional Gaussian distribution, and the width, height and rotation angle of the bounding box correspond to the covariance matrix of the two-dimensional Gaussian distribution, thus converting the rigid rotated bounding box into a continuous Gaussian probability distribution.

[0094] S1.4 Regression loss calculation based on Wasserstein distance: Based on the Wasserstein distance between two two-dimensional Gaussian distributions, a regression loss function for rotating bounding boxes is constructed; first, the Wasserstein distance between the Gaussian distribution corresponding to the predicted result and the Gaussian distribution corresponding to the true label is calculated, and then a smooth and continuous loss value is generated through nonlinear mapping.

[0095] S1.5 Iterative Model Training: Cross-entropy loss is used as the classification loss for the model's classification branch, and the loss constructed based on Wasserstein distance is used as the regression loss for the regression branch. These are combined to generate the model's total loss function. A stochastic gradient descent optimizer is used to iteratively update all the model's weight parameters through the backpropagation algorithm until the total loss function converges, thus completing the model's pre-training.

[0096] Furthermore, in another embodiment, S2 specifically includes the following:

[0097] S2.1 Hub Conveying and Station Positioning: The multi-station linear conveyor of the station positioning unit conveys the hub to be pressed to the servo rotary table 6 of the pressing station; the photoelectric sensor detects the hub's arrival signal and triggers the conveyor to stop; then the servo rotary table 6 is controlled to perform initial circumferential positioning of the hub, completing the fixing and locking of the hub at the pressing station;

[0098] S2.2 Image Acquisition Triggering and Lighting Control: After receiving the wheel hub positioning signal, the industrial controller synchronously triggers the surround industrial camera group 5 of the vision acquisition unit and the surface light source lighting system; it controls the surface light source lighting system to output a large area of ​​shadowless diffused light to uniformly illuminate the surface of the wheel hub.

[0099] S2.3 Synchronous Acquisition of Workpiece Images: Control the surround industrial camera group 5 to synchronously acquire color workpiece images containing all bolt holes of the wheel hub from multiple perspectives, and transmit the acquired image data to the industrial controller in real time;

[0100] S2.4 Workpiece Image Preprocessing: The industrial controller performs preprocessing operations on the acquired color workpiece images in sequence to generate images that meet the model input requirements. Specifically, this includes converting the color image into a single-channel grayscale image, using median filtering to denoise the grayscale image to suppress salt-and-pepper noise, enhancing the edge gradient information of the bolt holes through nonlinear grayscale transformation, and finally extracting the region of interest containing all bolt holes based on the outer contour features of the wheel hub.

[0101] Furthermore, in another embodiment, S3 specifically includes the following:

[0102] S3.1 Model Inference Output Rotation Bounding Box: Input the pre-processed workpiece image into the pre-trained deep learning model for rotating target detection, and output the core parameters of the rotation bounding box corresponding to each bolt hole in the image through model inference;

[0103] S3.2 Subpixel-level edge extraction: For each image region corresponding to a rotated bounding box, firstly, candidate edge pixels are extracted by gradient calculation, and then the position of the candidate edge pixels is optimized by subpixel-level interpolation through quadratic curve fitting to obtain more accurate edge coordinates;

[0104] S3.3 Ellipse Fitting and Hole Center Calculation: For the sub-pixel edge point set corresponding to each bolt hole, the least squares method is used to fit the ellipse model, and the coefficients of the ellipse model are obtained. Then, based on the fitted ellipse model, the center coordinates of the bolt hole are calculated.

[0105] S3.4 Topological Geometry Rigid Body Consistency Verification: First, obtain the number of bolt holes on the wheel hub to be press-fitted. Based on the geometric prior of the circumferentially uniformly distributed bolt holes, calculate the wheel hub geometric rotation center corresponding to the center coordinates of all bolt holes in sequence. Then, calculate the circumferential angle of each bolt hole relative to the wheel hub geometric rotation center based on the ellipse fitting result. If the circumferential angle deviation of a single bolt hole exceeds the preset threshold, it is judged as a false detection result and discarded. The wheel hub geometric rotation center and corresponding pose parameters are recalculated based on the fitting results of the remaining bolt holes.

[0106] S3.5 Bolt Hole Spatial Pose Output: The final output is the spatial pose parameters of each bolt hole, including the pixel coordinates of the bolt hole center, the circumferential angle relative to the geometric rotation center of the hub, and the hole diameter.

[0107] Furthermore, in another embodiment, S4 specifically includes the following:

[0108] S4.1 Hand-to-Hand Calibration Parameter Call: Calls the calibration parameters that have been determined in advance through the Eye-to-Hand calibration method, including the affine transformation relationship between the image pixel coordinate system and the physical coordinate system of the press-fitting actuator, and the circumferential angle mapping relationship between the image pixel coordinate system and the physical coordinate system of the servo rotary table 6.

[0109] S4.2 Pose Conversion and Compensation Calculation: Based on the affine transformation relationship obtained from calibration, the center pixel coordinates of the target bolt hole are converted into three-dimensional position coordinates in the physical coordinate system of the press-fitting actuator, and the position compensation of the press head is calculated; based on the circumferential angle mapping relationship obtained from calibration, the circumferential angle of the target bolt hole is converted into the rotation angle compensation of the servo rotary table 6.

[0110] S4.3 Coaxial Alignment Drive: Based on the calculated position compensation amount, the press head of the press actuator is driven to move directly above the target bolt hole; based on the calculated rotation angle compensation amount, the servo rotary table 6 is driven to rotate and adjust, so that the central axis of the press head is completely coincident with the central axis of the target bolt hole, thus completing the coaxial alignment.

[0111] Furthermore, in another embodiment, S5 specifically includes the following:

[0112] S5.1 Pressing Process Start-up and Real-time Signal Acquisition: The industrial controller starts the pressing process and controls the pressing head of the pressing actuator to move downward at a preset feed speed; the pressing force signal and displacement signal are acquired in real time at a preset sampling frequency through the pressure sensor and displacement sensor integrated at the end of the pressing head, and the visual pose signal of the pressing head and bolt hole is acquired in real time through the vision acquisition unit.

[0113] S5.2 Contact Stage Admittance Compliance Control: When the real-time pressing force exceeds the preset contact threshold, the system switches to admittance compliance control mode; the pressing head of the pressing actuator is equivalent to a second-order compliant system including virtual inertia, virtual damping, and virtual stiffness. A system dynamic model is established, and the deviation between the real-time collected pressing force and the preset guiding contact force is used as input. The real-time position correction of the pressing head is obtained through the dynamic model, and the feed position of the pressing head is dynamically adjusted until the bolt is detected to enter the preset guiding depth, thus ending this stage of control.

[0114] S5.3 Linear Active Disturbance Rejection Control during Pressing Stage: When the bolt is detected to have entered the preset guide depth, the system switches to linear active disturbance rejection control mode. First, the position correction amount of the admittance compliance control output is superimposed on the preset press-fit displacement trajectory to generate the final desired tracking trajectory. Then, a dynamic model of the press-fit actuator is established, and the nonlinear friction, dead zone, parameter fluctuations and external resistance disturbances inside the system are uniformly attributed to the total disturbance of the system. A third-order linear extended state observer is constructed to track and estimate the displacement state, velocity state and total disturbance of the system in real time. Based on the tracking error between the desired tracking trajectory and the actual state, combined with the real-time estimate of the total disturbance, the control quantity of the servo drive is calculated to control the feed action of the press head in real time and perform disturbance compensation until the press-fit displacement reaches the target depth and the press-fit force reaches the rated load, thus ending this stage of control.

[0115] S5.4 Constant Force Steady-State Control during the Holding Pressure Stage: When the pressing displacement reaches the target depth and the pressing force reaches the rated load, the system switches to constant force steady-state control mode. Based on the real-time collected deviation between the pressing force and the rated load, the position fine-tuning amount of the pressing head is calculated, and the position of the pressing head is dynamically adjusted. At the same time, through the linear expansion state observer, the equivalent disturbance caused by the elastic recovery of the material and the release of residual stress is estimated and compensated in real time to maintain the pressing force stable at the rated load. Until the holding pressure time reaches the preset value, the constant force steady-state control ends, and the pressing head is controlled to return at the preset speed to complete the entire bolt pressing process.

[0116] Furthermore, in one application example:

[0117] S1 pre-trained deep learning model for rotating object detection:

[0118] S1.1 Model Architecture Construction:

[0119] A single-stage anchor-free rotating object detection deep learning model is constructed. The backbone network adopts the CSPDarknet structure, and the feature fusion layer adopts the FPN+PAMet structure. The detection head is decoupled into a classification branch and a regression branch. The regression branch outputs five sets of parameters for the rotated bounding box, defined as follows: Five sets of parameters for the rotated bounding box: ; : The x-coordinate of the bounding box center in the image pixel coordinate system; : The ordinate of the bounding box center in the image pixel coordinate system; : Width of the bounding box; : Height of the bounding box; The rotation angle of the bounding box, with a value range of [value missing]. ;

[0120] S1.2 Training Sample Set Construction:

[0121] Image samples containing bolt holes of different wheel hub models were collected. Five sets of parameters of the corresponding rotated bounding box were labeled for the bolt holes in each sample to generate a labeled dataset. Data augmentation processing, including image normalization, random flipping, and brightness perturbation, was performed on the labeled dataset to construct a model training sample set.

[0122] S1.3 Two-dimensional Gaussian mapping of the rotated bounding box:

[0123] Rotate the bounding box predicted by the model. True rotated bounding boxes labeled in training samples These are mapped to two-dimensional Gaussian distributions, with the following mapping rules:

[0124] For five sets of parameters of an arbitrary rotated bounding box Mapped to a two-dimensional Gaussian distribution ,in:

[0125] Mean vector:

[0126] Covariance matrix:

[0127]

[0128] S1.4 Regression loss calculation based on Wasserstein distance:

[0129] Calculate the predicted Gaussian distribution Compared with the true Gaussian distribution The square of the two-dimensional Wasserstein distance between them :

[0130]

[0131] In the formula: For L2 norm operations, This is the trace operation of a matrix;

[0132] Calculate the regression loss of the rotated bounding box :

[0133]

[0134]

[0135] In the formula: This is a smoothing hyperparameter, with a value of 1 or 2;

[0136] S1.5 model iterative training:

[0137] Using cross-entropy loss as the classification loss, As the regression loss, the total loss function of the model is constructed; a stochastic gradient descent optimizer is used to iteratively update the model weights through backpropagation until the total loss function converges, thus completing the model training.

[0138] S2 wheel hub positioning and image acquisition:

[0139] S2.1 Hub Conveying and Station Positioning:

[0140] The multi-station linear conveyor of the station positioning unit transports the wheel hub to be pressed to the servo rotary table 6 of the pressing station; the photoelectric sensor detects the wheel hub's arrival signal and controls the conveyor to stop running; the servo rotary table 6 is controlled to perform circumferential initial positioning of the wheel hub, completing the fixed positioning of the wheel hub in the pressing station.

[0141] S2.2 Image Acquisition Triggering and Illumination Control:

[0142] After receiving the wheel hub positioning signal, the industrial controller synchronously triggers the surround industrial camera group 5 of the vision acquisition unit and the high-intensity surface light source illumination system; it controls the surface light source illumination system to output shadowless diffused light to illuminate the surface of the wheel hub.

[0143] S2.3 Synchronous acquisition of workpiece images:

[0144] The surround-type industrial camera group 5 is controlled to simultaneously acquire color workpiece images, including all bolt holes of the wheel hub, from multiple perspectives, and transmit the acquired image data to the industrial controller;

[0145] S2.4 Workpiece Image Preprocessing:

[0146] The industrial controller performs the following preprocessing operations sequentially on the acquired color workpiece images to generate preprocessed workpiece images:

[0147] 1. Grayscale conversion: Converting a color image into a single-channel grayscale image;

[0148] 2. Median filtering: A 3×3 window median filter is used to denoise the grayscale image and suppress salt-and-pepper noise;

[0149] 3. Gray-scale enhancement processing: Nonlinear gray-scale transformation is performed on the denoised image using gamma correction to enhance the gradient information at the bolt hole edges;

[0150] 4. Region of Interest Extraction: Based on the outer contour features of the wheel hub, extract the region of interest (RI) containing all bolt holes;

[0151] S3 calculates the spatial orientation of the wheel hub bolt holes:

[0152] S3.1 model inference outputs rotated bounding boxes:

[0153] The preprocessed workpiece image from S2 is input into the pre-trained deep learning model for rotating target detection from S1. The model infers and outputs five sets of parameters for the rotating bounding box corresponding to each bolt hole in the image. ;

[0154] S3.2 Subpixel-level Edge Extraction:

[0155] For each RI image region corresponding to a rotated bounding box, perform the following operations:

[0156] 1. The first-order gradient of the image is calculated using the Sbel operator to obtain the gradient magnitude. With gradient direction ;

[0157] 2. Extract candidate edge pixels along the gradient direction, and perform sub-pixel interpolation on the candidate edge pixels using quadratic curve fitting to calculate the sub-pixel coordinates of the edge points. The calculation formula is:

[0158]

[0159]

[0160] In the formula: The x and y coordinates of the edge points are at the pixel level. x-axis gradient magnitude at that point y-axis Gradient magnitude at;

[0161] S3.3 Ellipse Fitting and Hole Center Determination:

[0162] For each bolt hole, the sub-pixel edge point set is fitted with an ellipse using the least squares method to establish the ellipse model equation:

[0163]

[0164] Solving the equations yields the coefficients of the elliptic model. The center coordinates of the bolt hole were calculated based on the elliptical model. ;

[0165] S3.4 Topological Geometry Rigid Body Consistency Verification:

[0166] 1. Obtain the number of bolt holes for the wheel hub to be press-fitted. , Based on the circumferential direction of the hub bolt holes Using geometric priors that are evenly distributed at equal angles, calculate the hub geometric rotation center corresponding to the center coordinates of all bolt holes in sequence;

[0167] 2. For the ellipse fitting results of each bolt hole, calculate its circumferential angle relative to the geometric rotation center of the hub;

[0168] 3. If the circumferential angle deviation of a single bolt hole exceeds the preset angle threshold, the detection result of that bolt hole is discarded, and the spatial pose of the wheel hub geometric rotation center and each bolt hole is recalculated based on the fitting results of the remaining bolt holes.

[0169] S3.5 Bolt Hole Spatial Pose Output:

[0170] Output the spatial pose parameters of each bolt hole, including the pixel coordinates of the hole center, the circumferential angle relative to the geometric rotation center of the hub, and the hole diameter.

[0171] S4 actuator coaxial alignment:

[0172] S4.1 Hand-Eye Calibration Parameter Callup:

[0173] The calibration parameters, pre-determined using the Eye-t-Hand calibration method, are invoked, including the affine transformation matrix between the image pixel coordinate system and the physical coordinate system of the pressing actuator. The circumferential angle mapping relationship between the image pixel coordinate system and the physical coordinate system of the servo rotary table;

[0174] S4.2 Pose Transformation and Compensation Calculation:

[0175] 1. Based on affine transformation matrix The center pixel coordinates of the target bolt hole output by S3 are converted into three-dimensional position coordinates in the physical coordinate system of the press-fitting actuator, and the position compensation amount of the press head is calculated. ;

[0176] 2. Based on the circumferential angle mapping relationship, the circumferential angle of the target bolt hole is converted into the rotation angle compensation amount of the servo rotary table 6. ;

[0177] S4.3 coaxial alignment drive:

[0178] Based on location compensation amount The pressure head of the drive press-fitting actuator moves directly above the target bolt hole; based on the rotation angle compensation amount. Drive the servo rotary table 6 to rotate so that the central axis of the pressure head coincides with the central axis of the target bolt hole, thus completing the coaxial alignment;

[0179] S5 staged press-fit closed-loop control:

[0180] S5.1 Pressing Process Start-up and Real-time Signal Acquisition:

[0181] The industrial controller initiates the pressing process, controlling the pressing head of the pressing actuator to move downwards at a preset feed speed; the pressing force signal is collected in real time through pressure and displacement sensors integrated at the end of the pressing head at a preset sampling frequency. Displacement signal Simultaneously, the visual pose signals of the pressure head and bolt holes are acquired in real time through the visual acquisition unit; among which... Sampling time;

[0182] S5.2 Contact Phase Admittance Compliance Control:

[0183] When the pressing force signal is detected Exceeding the preset contact threshold When switching to admittance compliance control mode, perform the following operations:

[0184] 1. Establish the dynamic equations of the virtual mass-damping-stiffness second-order system of the pressure head:

[0185]

[0186] In the formula: For virtual inertia parameters, For virtual damping parameters, For virtual stiffness parameters, This is the position correction amount output by the admittance model. for The first derivative, for The second derivative, Preset guiding contact force;

[0187] 2. Based on the above dynamic equations, and according to the real-time force deviation... The position correction amount at the current sampling time is calculated. The output is sent to the press-fitting actuator to correct the feed position of the press head in real time;

[0188] 3. Until a displacement signal is detected. Once the preset guidance depth threshold is reached, the admittance compliance control mode ends.

[0189] S5.3 Linear Active Disturbance Rejection Control during the Push-in Stage:

[0190] When a displacement signal is detected When the preset guidance depth threshold is reached, switch to linear active disturbance rejection control mode and perform the following operations:

[0191] 1. Construct the desired tracking trajectory: Use the position correction amount output by S502 Superimposed on the preset press-fit displacement trajectory Generate the desired tracking trajectory The formula is:

[0192]

[0193] 2. Establish a dynamic model of the press-fitting actuator, and abstract the nonlinear dynamic model of the hydraulic / servo press-fitting system as follows:

[0194]

[0195] In the formula: The total disturbance function of the system includes internal nonlinear friction, dead zone, parameter perturbations, and external drag disturbances. External disturbances Given the control gain, This is the servo drive control voltage;

[0196] 3. Construct a third-order linear extended state observer (LES) to estimate the total system disturbance in real time. The equation of the LES is:

[0197]

[0198] In the formula: For the actual displacement of the system The tracking value, This is the tracked value of the system's actual speed. Total system disturbance Real-time estimates, For observer gain parameters;

[0199] 4. Calculate the tracking error and the final control variable:

[0200] Calculate position tracking error Speed ​​tracking error ;

[0201] PD control is used to calculate the initial control quantity. ,in For proportional gain, This is the differential gain;

[0202] Based on the total disturbance estimate The initial control input is compensated, and the final servo drive control voltage is calculated.

[0203]

[0204] 5. Control voltage The output is sent to the press-fitting actuator to control the press head to track the desired trajectory until a displacement signal is detected. Target depth reached and pressing force signal Reaching rated load The linear active disturbance rejection control mode ends;

[0205] S5.4 Constant force steady-state control during the pressure holding stage:

[0206] When a displacement signal is detected Target depth reached and pressing force signal Reaching rated load When switching to constant force steady-state control mode, perform the following operations:

[0207] 1. Calculate the pressing force error ;

[0208] 2. Based on pressing force error The position correction amount of the pressure head is calculated through the PI control algorithm, and the position of the pressure head is finely adjusted in real time.

[0209] 3. Simultaneously, a third-order linear expansion state observer is used to estimate and compensate for the equivalent disturbances caused by material elastic recovery and residual stress release in real time, maintaining the pressing force stable at the rated load. ;

[0210] 4. Continue holding the pressure until the preset holding time is reached. The constant force steady-state control mode ends, and the pressure head is controlled to return at a preset speed to complete the bolt pressing.

[0211] Please see the appendix Figure 4 The present invention provides an embodiment of a wheel hub bolt press-fitting control system based on deep learning, comprising a reinforced four-column reaction force bearing frame, a station positioning mechanism, a vision acquisition mechanism, a press-fitting execution mechanism, a feeding mechanism, and an industrial controller;

[0212] The reinforced four-column reaction load-bearing frame includes four vertically parallel precision guide columns 4, a top beam and a bottom base. The top beam is fixedly connected to the top of the four precision guide columns 4, and the bottom base is fixedly connected to the bottom of the four precision guide columns 4, forming a closed-loop load-bearing body.

[0213] The workstation positioning mechanism includes a multi-station linear conveyor and a servo rotary table 6. The multi-station linear conveyor is fixedly installed in the middle area of ​​the reinforced four-column reaction bearing frame, and the servo rotary table 6 is fixedly installed at the pressing station of the multi-station linear conveyor, with the workpiece bearing surface of the servo rotary table 6 facing upward.

[0214] The visual acquisition mechanism includes multiple industrial cameras and multiple sets of LED array panel lights 1. The multiple industrial cameras are symmetrically fixedly installed on the inner side walls of four precision guide columns 4, forming a camera group 5 surrounding the pressing station. The lenses of all industrial cameras face the workpiece bearing surface of the servo rotary table 6. The multiple sets of LED array panel lights 1 are fixedly installed on the lower surface of the top crossbeam, and the light-emitting surfaces of all LED array panel lights 1 face the workpiece bearing surface of the servo rotary table 6.

[0215] The press-fitting actuator includes a dual servo hydraulic power source 2 and a press head. The dual servo hydraulic power source 2 is fixedly installed at the center of the top crossbeam. The power output end of the dual servo hydraulic power source 2 is fixedly connected to the press head through a precision guide rod. The press head is located above the servo rotary table 6. The lower end of the press head integrates a pressure sensor and a displacement sensor.

[0216] The feeding mechanism is an industrial robot 3, which is set on the feeding side of the reinforced four-column reaction load-bearing frame; the industrial controller is electrically connected to the multi-station linear conveyor, servo rotary table 6, industrial camera, LED array panel light 1, dual servo hydraulic 2 power source, industrial robot 3, pressure sensor, and displacement sensor respectively.

[0217] The industrial controller is configured to execute a deep learning-based wheel hub bolt press-fit control method.

[0218] Furthermore, in another application example, multiple sets of comparative tests were conducted on a mixed-line production line for automotive parts. The experiments compared and verified the existing traditional technologies from three dimensions: visual algorithm positioning performance, compliant control anti-interference performance, and overall system pressing efficiency.

[0219] Experiment 1: Comparison of Positioning Accuracy and Real-Time Performance of Visual Algorithm for Bolt Holes

[0220] Experimental objective: To verify the effectiveness of this invention in solving the problem of angular jumps in near-circular targets and improving detection speed and accuracy;

[0221] Experimental subjects: 1000 randomly selected images of wheel hub bolt holes with complex backgrounds (covering different lighting and multi-view distortion).

[0222] Comparison of options:

[0223] Solution A (existing technology): Faster R-CNN two-stage detection algorithm + traditional Canny edge detection;

[0224] Option B (Conventional Single-Stage): Standard YOLOv8-OBB + Smooth-L1 loss function;

[0225] Experimental data results:

[0226]

[0227] Results: Data shows that after introducing the GWD loss function, this invention eliminates the problem of drastic angle jumps caused by near-circular targets in square B, increasing the detection success rate to 99.6%. Combined with sub-pixel interpolation, while maintaining millisecond-level (21ms) ultra-fast inference, the positioning accuracy reaches the micrometer level (0.1 pixel level), providing accurate feedforward input for subsequent mechanical actions.

[0228] Experiment 2: Comparison of rigid collision resistance and compliance under admittance-active disturbance rejection composite control:

[0229] Experimental objective: To verify the effectiveness of the proposed dual closed-loop architecture of "outer loop admittance compliant control - inner loop self-disturbance rejection robust tracking control" in reducing the instantaneous peak force during the initial contact stage of press fitting, reducing the incidence of damage such as thread scratches and stripping, and improving trajectory tracking accuracy and press fitting stability under nonlinear disturbances, under the condition of small position deviations and angle skews.

[0230] Experimental setup: A slight positional deviation of 0.5mm~2.0mm and an angular deviation of 1°~3° were artificially set between the center of the pressure head and the center of the bolt hole to simulate the cumulative error under visual blind spot conditions, and 500 pressing tests were performed respectively.

[0231] Comparison of options:

[0232] Option A (Traditional Rigid Press-fit): Pure Position PID Control;

[0233] Option B (Traditional Single Compliance Control): Employs constant stiffness impedance control;

[0234] The present invention provides a composite compliant control system based on admittance and active disturbance rejection.

[0235]

[0236] Results: Experimental results show that: Scheme A generates an instantaneous contact force peak exceeding 850N during misaligned contact and lacks a yielding mechanism. In 500 tests, thread scratches or stripping occurred 41 times, with a damage rate of 8.2% and a trajectory tracking error of ±0.5mm. Scheme B has an instantaneous contact force peak of approximately 320N, a resistance yielding response time of approximately 85ms, and in 500 tests, thread scratches or stripping occurred 13 times, with a damage rate of 2.6% and a trajectory tracking error of ±0.3mm. The present invention can control the instantaneous contact force peak to within 45N. Under these conditions, the response time for obstructed retraction is less than 10ms. No thread scratches, stripping, or jamming occurred in 500 tests, with a damage rate of 0%. The trajectory tracking error is controlled within ±0.02mm. This indicates that the present invention can significantly reduce the peak contact force caused by rigid collisions, improve the dynamic response speed of obstructed retraction, and effectively suppress failures such as thread scratches, stripping, and jamming, while ensuring high-precision trajectory tracking during the press-fitting process. This demonstrates that the present invention has excellent compliance, anti-interference ability, and press-fitting safety.

[0237] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A deep learning-based control method for wheel hub bolt press-fitting, characterized in that, Includes the following steps: S1 Pre-trained Rotating Target Detection Deep Learning Model: Construct a rotating target detection deep learning model for wheel hub bolt hole detection. Convert the bolt hole annotations in the training samples into a two-dimensional Gaussian distribution. Calculate the regression loss of the rotating bounding box using the Wasserstein distance between the two two-dimensional Gaussian distributions to complete model training. S2 Wheel Hub Positioning and Image Acquisition: The wheel hub to be pressed is transported and positioned to the pressing station through the station positioning unit, and the workpiece image including the wheel hub bolt holes is acquired through the vision acquisition unit; S3 calculates the spatial pose of the wheel hub bolt holes: The acquired workpiece image is input into the model pre-trained by S1, and the rotated bounding box of the bolt hole is output; Sub-pixel level edge extraction and ellipse fitting are performed on the rotated bounding box region, and combined with the geometric prior of the circumferentially distributed wheel hub bolt holes, the topological geometric rigid body consistency of the fitting result is checked, and the spatial pose of the bolt hole is calculated. S4 Actuator Coaxial Alignment: Based on the affine transformation matrix between the image pixel coordinate system determined in advance through eye-to-hand calibration and the physical coordinate system of the press-fitting actuator, the spatial pose of the target bolt hole obtained in S3 is converted into position and angle compensation amounts, driving the press head of the press-fitting actuator to be coaxially aligned with the target bolt hole. S5 phased press-fit closed-loop control: Start the press-fit process and collect the press-fit force and displacement of the press-fit actuator and the real-time visual pose signal of the vision acquisition unit in real time; When the press contact force is detected to exceed the preset contact threshold, the system switches to admittance compliance control mode, which treats the press head as a virtual mass-damping-stiffness second-order system and generates a position correction amount based on the real-time force deviation. When the bolt is detected to have entered the preset guide depth, the system switches to the linear active disturbance rejection control mode, and the position correction amount is superimposed on the preset press-fit displacement trajectory to generate the desired tracking trajectory and complete the tracking. When the pressing displacement reaches the target depth and the pressing force reaches the rated load, switch to constant force steady-state control mode, maintain the preset constant force until the pressure holding is completed, and end the bolt pressing.

2. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: S1 specifically includes the following: S1.1 Model Architecture Construction: A single-stage anchor-free rotating object detection deep learning model is constructed, which consists of three parts: backbone network, feature fusion layer, and decoupled detection head; the backbone network adopts the CSPDarknet structure to extract multi-scale features of the input image; The feature fusion layer adopts the FPN+PAMet structure to achieve efficient fusion of deep semantic features and shallow detail features of the image; the decoupled detection head is divided into independent classification branches and regression branches. The regression branch is used to output five sets of core parameters of the bolt hole rotation bounding box, namely the pixel coordinates of the bounding box center, the width and height of the bounding box, and the rotation angle of the bounding box. S1.2 Construction of Training Sample Set: Collect image samples covering bolt holes of different wheel hub models, label the core parameters of the corresponding rotated bounding box for each bolt hole in the sample, and generate a labeled dataset; Data augmentation processes such as image normalization, random flipping, and brightness perturbation are performed on the labeled dataset to enrich the sample richness and finally construct the model training sample set. S1.3 Two-dimensional Gaussian mapping of rotated bounding boxes: The rotated bounding boxes predicted by the model and the real rotated bounding boxes labeled in the training samples are mapped to two-dimensional Gaussian distributions respectively; where the center coordinates of the rotated bounding box correspond to the mean vector of the two-dimensional Gaussian distribution, and the width, height and rotation angle of the bounding box correspond to the covariance matrix of the two-dimensional Gaussian distribution, thus converting the rigid rotated bounding box into a continuous Gaussian probability distribution. S1.4 Regression loss calculation based on Wasserstein distance: Based on the Wasserstein distance between two two-dimensional Gaussian distributions, a regression loss function for rotating bounding boxes is constructed; first, the Wasserstein distance between the Gaussian distribution corresponding to the predicted result and the Gaussian distribution corresponding to the true label is calculated, and then a smooth and continuous loss value is generated through nonlinear mapping. S1.5 Iterative Model Training: Cross-entropy loss is used as the classification loss for the model's classification branch, and the loss constructed based on Wasserstein distance is used as the regression loss for the regression branch. These are combined to generate the model's total loss function. A stochastic gradient descent optimizer is used to iteratively update all the model's weight parameters through the backpropagation algorithm until the total loss function converges, thus completing the model's pre-training.

3. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: S2 specifically includes the following: S2.1 Hub Conveying and Station Positioning: The hub to be pressed is conveyed to the servo rotary table (6) of the pressing station through the multi-station linear conveyor of the station positioning unit; the hub arrival signal is detected by the photoelectric sensor, triggering the conveyor to stop running; then the servo rotary table (6) is controlled to perform circumferential initial positioning of the hub, and the hub is fixed and locked in the pressing station. S2.2 Image Acquisition Trigger and Lighting Control: After receiving the wheel hub positioning signal, the industrial controller synchronously triggers the surround industrial camera group (5) of the vision acquisition unit and the surface light source lighting system; controls the surface light source lighting system to output a large area of ​​shadowless diffused light to uniformly illuminate the surface of the wheel hub; S2.3 Synchronous acquisition of workpiece images: Control the surround industrial camera group (5) to synchronously acquire color workpiece images containing all bolt holes of the hub from multiple perspectives, and transmit the acquired image data to the industrial controller in real time; S2.4 Workpiece Image Preprocessing: The industrial controller performs preprocessing operations on the acquired color workpiece images in sequence to generate images that meet the model input requirements. Specifically, this includes converting the color image into a single-channel grayscale image, using median filtering to denoise the grayscale image to suppress salt-and-pepper noise, enhancing the edge gradient information of the bolt holes through nonlinear grayscale transformation, and finally extracting the region of interest containing all bolt holes based on the outer contour features of the wheel hub.

4. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: S3 specifically includes the following: S3.1 Model Inference Output Rotation Bounding Box: Input the pre-processed workpiece image into the pre-trained deep learning model for rotating target detection, and output the core parameters of the rotation bounding box corresponding to each bolt hole in the image through model inference; S3.2 Subpixel-level edge extraction: For each image region corresponding to a rotated bounding box, firstly, candidate edge pixels are extracted by gradient calculation, and then the position of the candidate edge pixels is optimized by subpixel-level interpolation through quadratic curve fitting to obtain more accurate edge coordinates; S3.3 Ellipse Fitting and Hole Center Calculation: For the sub-pixel edge point set corresponding to each bolt hole, the least squares method is used to fit the ellipse model, and the coefficients of the ellipse model are obtained. Then, based on the fitted ellipse model, the center coordinates of the bolt hole are calculated. S3.4 Topological Geometry Rigid Body Consistency Verification: First, obtain the number of bolt holes on the wheel hub to be press-fitted. Based on the geometric prior of the circumferentially uniformly distributed bolt holes, calculate the wheel hub geometric rotation center corresponding to the center coordinates of all bolt holes in sequence. Then, calculate the circumferential angle of each bolt hole relative to the wheel hub geometric rotation center based on the ellipse fitting result. If the circumferential angle deviation of a single bolt hole exceeds the preset threshold, it is judged as a false detection result and discarded. The wheel hub geometric rotation center and corresponding pose parameters are recalculated based on the fitting results of the remaining bolt holes. S3.5 Bolt Hole Spatial Pose Output: The final output is the spatial pose parameters of each bolt hole, including the pixel coordinates of the bolt hole center, the circumferential angle relative to the geometric rotation center of the hub, and the hole diameter.

5. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: S4 specifically includes the following: S4.1 Hand-to-Hand calibration parameter call: Call the calibration parameters determined in advance by the Eye-to-Hand calibration method, including the affine transformation relationship between the image pixel coordinate system and the physical coordinate system of the press-fitting actuator, and the circumferential angle mapping relationship between the image pixel coordinate system and the physical coordinate system of the servo rotary table (6); S4.2 Pose Conversion and Compensation Calculation: Based on the affine transformation relationship obtained by calibration, the center pixel coordinates of the target bolt hole are converted into three-dimensional position coordinates under the physical coordinate system of the press actuator, and the position compensation of the press head is calculated; Based on the circumferential angle mapping relationship obtained by calibration, the circumferential angle of the target bolt hole is converted into the rotation angle compensation of the servo rotary table (6); S4.3 Coaxial alignment drive: Based on the calculated position compensation amount, drive the press head of the press actuator to move directly above the target bolt hole; based on the calculated rotation angle compensation amount, drive the servo rotary table (6) to rotate and adjust, so that the central axis of the press head is completely coincident with the central axis of the target bolt hole, and complete the coaxial alignment.

6. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: S5 specifically includes the following: S5.1 Pressing Process Start-up and Real-time Signal Acquisition: The industrial controller starts the pressing process and controls the pressing head of the pressing actuator to move downward at a preset feed speed; the pressing force signal and displacement signal are acquired in real time at a preset sampling frequency through the pressure sensor and displacement sensor integrated at the end of the pressing head, and the visual pose signal of the pressing head and bolt hole is acquired in real time through the vision acquisition unit. S5.2 Contact Stage Admittance Compliance Control: When the real-time pressing force exceeds the preset contact threshold, the system switches to admittance compliance control mode; the pressing head of the pressing actuator is equivalent to a second-order compliant system including virtual inertia, virtual damping, and virtual stiffness. A system dynamic model is established, and the deviation between the real-time collected pressing force and the preset guiding contact force is used as input. The real-time position correction of the pressing head is obtained through the dynamic model, and the feed position of the pressing head is dynamically adjusted until the bolt is detected to enter the preset guiding depth, thus ending this stage of control. S5.3 Linear Active Disturbance Rejection Control during Pressing Stage: When the bolt is detected to have entered the preset guide depth, the system switches to linear active disturbance rejection control mode. First, the position correction amount of the admittance compliance control output is superimposed on the preset press-fit displacement trajectory to generate the final desired tracking trajectory. Then, a dynamic model of the press-fit actuator is established, and the nonlinear friction, dead zone, parameter fluctuations and external resistance disturbances inside the system are uniformly attributed to the total disturbance of the system. A third-order linear extended state observer is constructed to track and estimate the displacement state, velocity state and total disturbance of the system in real time. Based on the tracking error between the desired tracking trajectory and the actual state, combined with the real-time estimate of the total disturbance, the control quantity of the servo drive is calculated to control the feed action of the press head in real time and perform disturbance compensation until the press-fit displacement reaches the target depth and the press-fit force reaches the rated load, thus ending this stage of control. S5.4 Constant Force Steady-State Control during the Holding Pressure Stage: When the pressing displacement reaches the target depth and the pressing force reaches the rated load, the system switches to constant force steady-state control mode. Based on the real-time collected deviation between the pressing force and the rated load, the position fine-tuning amount of the pressing head is calculated, and the position of the pressing head is dynamically adjusted. At the same time, through the linear expansion state observer, the equivalent disturbance caused by the elastic recovery of the material and the release of residual stress is estimated and compensated in real time to maintain the pressing force stable at the rated load. Until the holding pressure time reaches the preset value, the constant force steady-state control ends, and the pressing head is controlled to return at the preset speed to complete the entire bolt pressing process.

7. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: The specific steps for converting the rotated bounding box into a two-dimensional Gaussian distribution are as follows: For five sets of parameters of an arbitrary rotated bounding box Mapped to a two-dimensional Gaussian distribution ,in: Mean vector: Covariance matrix: In the formula: This is the x-coordinate of the bounding box center in the image pixel coordinate system; The ordinate of the bounding box center in the image pixel coordinate system; This represents the width of the bounding box; This represents the height of the bounding box; The rotation angle of the bounding box, with a value range of 1. .

8. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: The regression loss of the rotated bounding box is calculated as follows: Calculate the predicted Gaussian distribution Compared with the true Gaussian distribution The square of the two-dimensional Wasserstein distance between them : In the formula: For L2 norm operations; This is the trace operation of a matrix; The rotated bounding box predicted by the model; The actual rotated bounding boxes labeled in the training samples; Calculate the regression loss of the rotated bounding box : In the formula: For smoothing hyperparameters.

9. The wheel hub bolt press-fitting control method based on deep learning according to claim 1, characterized in that: When the pressing force signal is detected Exceeding the preset contact threshold When switching to admittance compliance control mode, perform the following operations: Establish the second-order dynamic equations for the virtual mass-damping-stiffness system of the pressure head: In the formula: For virtual inertia parameters, For virtual damping parameters, For virtual stiffness parameters, This is the position correction amount output by the admittance model. for The first derivative, for The second derivative, Preset guiding contact force. Based on the above dynamic equations, according to the real-time force deviation The position correction amount at the current sampling time is calculated. The output is sent to the press-fitting actuator to correct the feed position of the press head in real time; Until a displacement signal is detected Once the preset guidance depth threshold is reached, the admittance compliance control mode ends.

10. A wheel hub bolt press-fitting control system based on deep learning, characterized in that: It includes a reinforced four-column reaction load-bearing frame, a workstation positioning mechanism, a vision acquisition mechanism, a pressing execution mechanism, a feeding mechanism, and an industrial controller; The reinforced four-column reaction bearing frame includes four vertically parallel precision guide columns (4), a top beam and a bottom base. The top beam is fixedly connected to the top of the four precision guide columns (4), and the bottom base is fixedly connected to the bottom of the four precision guide columns (4), forming a closed-loop bearing body. The station positioning mechanism includes a multi-station linear conveyor and a servo rotary table (6). The multi-station linear conveyor is fixedly installed in the middle area of ​​the reinforced four-column reaction bearing frame. The servo rotary table (6) is fixedly installed at the pressing station of the multi-station linear conveyor, and the workpiece bearing surface of the servo rotary table (6) is set upward. The vision acquisition mechanism includes multiple industrial cameras and multiple sets of LED array panel lights (1). The multiple industrial cameras are symmetrically fixedly installed on the inner sidewalls of four precision guide columns (4) to form a camera group (5) surrounding the press-fitting station. The lenses of all industrial cameras face the workpiece bearing surface of the servo rotary table (6). The multiple sets of LED array panel lights (1) are fixedly installed on the lower surface of the top crossbeam. The light-emitting surfaces of all LED array panel lights (1) face the workpiece bearing surface of the servo rotary table (6). The press-fitting actuator includes a dual servo hydraulic (2) power source and a press head. The dual servo hydraulic (2) power source is fixedly installed at the center of the top crossbeam. The power output end of the dual servo hydraulic (2) power source is fixedly connected to the press head through a precision guide rod. The press head is located above the servo rotary table (6). The lower end of the press head integrates a pressure sensor and a displacement sensor. The feeding mechanism is an industrial robot (3), which is set on the feeding side of the reinforced four-column reaction bearing frame; the industrial controller is electrically connected to the multi-station linear conveyor, servo rotary table (6), industrial camera, LED array panel light (1), dual servo hydraulic (2) power source, industrial robot (3), pressure sensor and displacement sensor respectively. The industrial controller is configured to execute the deep learning-based wheel hub bolt press-fitting control method as described in claim 1.

Images