Constant force polishing dynamic parameter self-tuning method
Patent Information
- Application Number
- CN202610933010.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-26
- Publication Date
- 2026-08-18
AI Technical Summary
[0003]参数固化，适应性差：控制器参数（如PID增益、阻抗质量-阻尼-刚度参数）通常为固定值，在出厂或现场调试时设定
[0034] 1. From passive response to active adaptation: Through visual foresight, the system can anticipate changes in operating conditions and adjust parameters in advance, greatly reducing force overshoot and oscillation, and achieving a smooth transition.
Smart Images
Figure CN122584080A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control technology for industrial robots, specifically a method for self-tuning dynamic parameters of constant force grinding. Background Technology
[0002] Robotic constant-force grinding is a key process in automated manufacturing, its core being maintaining a constant contact force between the tool tip and the workpiece. Currently, mainstream technologies employ methods based on impedance control or PID control. However, these methods have inherent drawbacks:
[0003] Fixed parameters and poor adaptability: Controller parameters (such as PID gain, impedance-mass-damping-stiffness parameters) are usually fixed values, set at the factory or during field commissioning. When the grinding object changes from one material to another, or when the workpiece geometry changes abruptly, fixed parameters cannot guarantee optimal control performance, easily leading to force overshoot, under-grinding, or system oscillation.
[0004] Response lag: Traditional control is a "post-event compensation", that is, it only adjusts after the force error occurs, and cannot cope with dynamically changing working conditions.
[0005] Reliance on expert experience: Parameter optimization relies heavily on the experience of engineers and repeated trial and error, resulting in long debugging cycles and high costs.
[0006] While existing technologies have improved calibration accuracy, they have not solved the problem of online parameter adaptation. Therefore, there is an urgent need for an intelligent system capable of sensing changes in operating conditions and autonomously adjusting control parameters. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a dynamic parameter self-tuning method that integrates visual forward perception and deep learning decision-making. This method can predict changes in working conditions in advance and adjust force control parameters in real time, thereby achieving truly adaptive constant force grinding.
[0008] The technical solution adopted by the present invention to achieve the above objectives is as follows:
[0009] A method for self-tuning dynamic parameters of constant force grinding includes the following steps:
[0010] 1) When the robot moves along the predetermined trajectory, multimodal sensor data is collected in real time, and data features are extracted;
[0011] 2) Using the extracted multimodal feature vectors as input, the parameters of the underlying robot force controller are predicted through a pre-trained deep learning model;
[0012] 3) The robot force controller uses new parameters and contact force feedback signals to drive the industrial robot to perform grinding operations through impedance control algorithms and generate joint torque commands.
[0013] The multimodal sensing data includes: 3D point cloud data of the workpiece surface in front of the robot's motion path, torque data between the robot's end flange and the constant force actuator, three-axis acceleration data of the actuator body, and grinding audio data.
[0014] The extracted data features are specifically as follows:
[0015] Visual feature extraction: For point cloud data, calculate the average curvature C and the gradient of the normal vector change of the path segment in front of the center point of the robot tool with a set distance and length;
[0016] Physical feature extraction: Calculate the RMS value A of the acceleration signal over a set time period in the past. rms and the frequency band energy A at the set frequency freq And calculate the Mel-frequency cepstral coefficient characteristics M1-M13 of the audio signal within the same time period;
[0017] Feature fusion: combining C, G, and A rms A freq The 17 eigenvalues M1-M13 are concatenated to form an eigenvector V. t .
[0018] Step 2) specifically refers to:
[0019] The feature vector V of the current time and the past 9 time points. t V t-1 , ..., V t-9 A 10×17 dimensional matrix is constructed in chronological order and used as the input to an LSTM network. After training, the LSTM network outputs the optimal stiffness parameter K. and damping parameter B .
[0020] Step 3) specifically refers to:
[0021] The robot force controller performs low-pass filtering, zero-point calibration, and tool gravity compensation on the actual contact force collected by the torque sensor and the position and speed feedback signals of the joint encoder to obtain the real workpiece contact force.
[0022] The actual pose of the robot's end effector is solved by forward kinematics, and the coordinate system reference of all computational quantities is unified.
[0023] By substituting the deviation between the desired trajectory and the actual pose, as well as the contact force deviation, into the Cartesian space impedance control model through impedance control calculation, and combining the current stiffness and damping parameters, the required correction force of the robot end effector is calculated.
[0024] When the actual contact force is greater than the target value, the model generates a corrective force in the yielding direction, driving the end effector to retreat accordingly; when the contact force is less than the target value, it generates a corrective force in the feeding direction, pushing the end effector to apply pressure compensation.
[0025] The robot force controller synchronously calculates the feedforward torque through the inverse dynamics model, and then converts the modified force in Cartesian space into the compensation torque of each joint through the force mapping relationship of the Jacobian matrix. The torque is then superimposed with the feedforward torque to generate the final joint torque command, which is sent to the joint servo driver for execution via the real-time bus.
[0026] It also includes an online adaptive optimization step, specifically:
[0027] Record the force control performance index after each force control parameter adjustment, periodically use a batch of new sensor data sequences and their corresponding force control performance indexes as incremental training samples to incrementally learn the deep learning model, update the model weights, load the updated model weights into the robot force controller during the robot's idle period, and switch to the new model before the next workpiece grinding begins.
[0028] A constant force grinding dynamic parameter self-tuning system includes:
[0029] The feature acquisition and extraction module is used to acquire multimodal sensor data in real time and extract data features when the robot moves along a predetermined trajectory;
[0030] The deep learning decision module is used to predict the parameters of the underlying robot force controller by taking the extracted multimodal feature vectors as input and using a pre-trained deep learning model.
[0031] The robot force controller module is used by the robot force controller to use new parameters and contact force feedback signals, through impedance control algorithms, to generate joint torque commands and drive the industrial robot to perform grinding operations.
[0032] It also includes a reinforcement learning module, which records the force control performance index after each force control parameter adjustment, periodically uses a batch of new sensor data sequences and their corresponding force control performance indexes as incremental training samples to incrementally learn the deep learning model, update the model weights, load the updated model weights into the robot force controller during the robot's idle period, and switch to the new model before the next workpiece grinding begins.
[0033] The present invention has the following beneficial effects and advantages:
[0034] 1. From passive response to active adaptation: Through visual foresight, the system can anticipate changes in operating conditions and adjust parameters in advance, greatly reducing force overshoot and oscillation, and achieving a smooth transition.
[0035] 2. Significantly improves the consistency of grinding quality: Faced with changes in workpiece geometry and material, the system can automatically maintain the optimal force control state, ensuring that the grinding effect is uniform and consistent throughout the complex curved workpiece.
[0036] 3. Reduced debugging and maintenance costs: Eliminates the process of repeatedly manually adjusting parameters for different workpieces, reduces the skill requirements for operators, and improves production efficiency.
[0037] 4. Possesses continuous evolution potential: Through online learning capabilities, the system can accumulate experience with use, continuously optimize performance, and possess the typical characteristics of industrial AI. Attached Figure Description
[0038] Figure 1 System framework diagram. Detailed Implementation
[0039] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0040] like Figure 1 As shown, a constant force grinding dynamic parameter self-tuning system consists of the following parts:
[0041] Robotic polishing platform: includes a visual perception module and a physical sensing module.
[0042] The visual perception module consists of a structured light 3D camera fixed in front of the robot's end effector, used to acquire real-time 3D point cloud data of the workpiece surface in front of the robot's motion path. Its field of view covers at least 0.5 seconds of the robot's motion trajectory.
[0043] The physical sensing module includes a torque sensor installed between the robot's end flange and the constant force actuator, a three-axis accelerometer installed on the actuator body, and a microphone for collecting grinding sounds.
[0044] Computation and control platform: including feature extraction and fusion module, deep learning decision module, dynamic parameter tuning and execution module, robot force controller module, and reinforcement learning module.
[0045] Feature extraction and fusion module: Connected to each sensor, it processes 3D point cloud data, calculates the average curvature and normal vector gradient of the area to be polished; extracts the root mean square (RMS) value and main frequency band energy of vibration energy from accelerometer signals; and extracts the first 13 coefficients of Mel-frequency cepstral coefficients (MFCC) from microphone signals as acoustic features. The above visual and physical features are then concatenated into a unified multimodal feature vector.
[0046] The deep learning decision-making module's core is a pre-trained Long Short-Term Memory (LSTM) network. The input to this module is the aforementioned multimodal feature vector, and the output is the optimal stiffness parameter K of the underlying robot force controller (using an impedance control model). and damping parameter B .
[0047] Dynamic parameter tuning execution module: Receives parameters output from the LSTM network and updates the impedance control parameters in the robot force controller in real time.
[0048] Robot force controller module: It has a built-in impedance control model, receives the real-time stiffness K and damping B output by the dynamic parameter tuning execution module, combines the contact force feedback signal collected by the torque sensor, executes the impedance control algorithm and generates joint torque commands, and drives the industrial robot body to complete high-precision, constant-force stable grinding operations according to the planned trajectory.
[0049] The robot force controller module incorporates a Cartesian space impedance control model. This model's algorithm operates in a closed loop with a 1kHz control cycle. Its core principle is to simulate compliant mechanical properties at the robot's end effector, enabling adaptive adjustment of the contact state upon workpiece contact for stable, constant-force grinding. The controller receives stiffness and damping parameters in real-time from the dynamic parameter tuning execution module, simultaneously acquiring the actual contact force from the torque sensor, the position and velocity feedback from the joint encoder, and the desired motion trajectory output from the upper-level trajectory planning layer. During operation, the sensor signals are first subjected to low-pass filtering, zero-point calibration, and tool gravity compensation to eliminate high-frequency vibration noise and tool weight interference, obtaining the true workpiece contact force. Simultaneously, the actual end effector pose is solved using forward kinematics, unifying the coordinate system reference for all computational quantities. Based on this, impedance control calculations are performed, substituting the deviation between the desired trajectory and the actual pose, as well as the contact force deviation, into the impedance model. Combined with the current stiffness and damping parameters, the required correction force for the end effector is calculated. When the actual contact force exceeds the target value, the model generates a corrective force in the yielding direction, driving the end effector to retreat accordingly. When the contact force is less than the target value, a corrective force in the feed direction is generated, pushing the end effector to apply pressure compensation, thereby achieving closed-loop regulation of the contact force. To counteract the interference of the robot's own weight, inertia, and Coriolis force, the controller synchronously calculates the feedforward torque through the inverse dynamics model; then, through the force mapping relationship of the Jacobian matrix, it converts the corrective force in Cartesian space into compensation torque for each joint, which is superimposed with the feedforward torque to generate the final joint torque command, which is sent to the joint servo driver for execution via the real-time bus.
[0050] The reinforcement learning module records the force control performance indicators after each parameter adjustment, such as the integral of the absolute value of the force error (IAE). Periodically (e.g., after each workpiece grinding is completed), a new batch of data (feature vector sequence, performance indicators) is sent to the reinforcement learning module. The reinforcement learning module uses this new data to incrementally learn (fine-tune) the LSTM model. This module automatically returns the learned data to the robot force controller module during the robot's next idle period (e.g., during part changeovers, before the end of the shift). Upon receiving the new model, the robot force controller module first stores it in non-volatile memory. Without interrupting the current grinding operation, it loads and initializes the model weights in the background. Before the next workpiece grinding begins, it automatically switches to the new version of the LSTM model, enabling the model to adapt to new tool wear conditions or new materials.
[0051] The process of a method for self-tuning dynamic parameters of constant force grinding is as follows:
[0052] S1: Real-time data acquisition and feature extraction:
[0053] As the robot moves along the predetermined trajectory, the visual perception module continuously collects 3D point cloud data of the workpiece surface in front, while the physical sensing module simultaneously collects six-dimensional torque, three-axis acceleration, and audio data.
[0054] The feature extraction and fusion module performs the following operations at a frequency of 1kHz:
[0055] Visual feature extraction: For the latest point cloud data, calculate the average curvature C and the gradient of the normal vector change of a path segment with a length of 20 mm located 50 mm in front of the robot tool center point (TCP).
[0056] Physical feature extraction: Calculate the RMS value (A) of the acceleration signal over the past 0.1 seconds. rms ) and 100-500Hz frequency band energy (A freq ); Calculate the MFCC features (M1-M13) of the audio signal within the same time period.
[0057] Feature fusion: combining curvature C, gradient G, and A rms A freq The 17 eigenvalues M1-M13 are concatenated to form an eigenvector V. t .
[0058] S2 Parameter Prediction Based on Deep Learning:
[0059] The feature vectors (V) of the current time and the past 9 times. t V t-1 , ..., V t-9The data are arranged in chronological order to form a 10×17 dimensional matrix, which serves as the input to the LSTM network. The LSTM network remembers historical operating conditions through its internal states.
[0060] The LSTM network is trained to output two floating-point numbers: a stiffness parameter K and a damping parameter B. This network can learn complex mappings such as "in regions of high curvature, stiffness needs to be reduced and damping increased to prevent oscillations."
[0061] S3 dynamic parameter issuance and execution:
[0062] The dynamic parameter tuning execution module receives (K) B After that, smoothly set it as the current parameter of the robot impedance controller.
[0063] The robot force controller uses new parameters to execute force control cycles, achieving precise control of contact forces.
[0064] S4: Online Adaptive Optimization
[0065] The system records force control performance indicators after each parameter adjustment, such as the integral of the absolute value of the force error (IAE). Periodically (e.g., after each workpiece grinding is completed), a new batch of data (feature vector sequence, performance indicators) is sent to the robot force controller module. The robot force controller module uses this new data to incrementally learn (fine-tune) the LSTM model, enabling the model to adapt to new tool wear conditions or new materials.
[0066] Example
[0067] A constant force actuator with a rated working air pressure of 0.6MPa and a theoretical output force range of 0-150N is mounted on a six-axis industrial robotic arm. The grinding constant force is set to 50N. This is used as an example to grind an aerospace aluminum alloy workpiece containing a planar area and a boss transition area with a height of 8mm. The self-tuning method for dynamic parameters of constant force grinding is described in detail. The system grinding robot platform includes a vision perception module (3D camera), a physical sensor module (force sensor, three-axis accelerometer, industrial microphone), and a computing and control platform.
[0068] Real-time step-by-step details:
[0069] Phase 1, Detailed Explanation of Deep Learning Model Training:
[0070] Training data acquisition process: On test workpieces with 5 different geometric features (plane, boss, groove, ramp, curved surface), experienced engineers manually calibrated the system. For each working condition, the engineers repeatedly adjusted the impedance control parameters (stiffness K and damping B) until the force control effect was optimal (force error < ±1N, smooth transition without oscillation). The system synchronously recorded the optimal parameter combination and the corresponding multimodal sensor data, collecting a total of 12,000 sets of valid "working condition-parameter" mapping data.
[0071] Model training parameters: The network structure is a 2-layer bidirectional LSTM with 128 hidden units per layer and a dropout rate of 0.2; input features are a 17-dimensional feature vector, including: curvature and normal gradient G from vision, vibration RMS value and 100-500Hz frequency band energy from the accelerometer; 13-dimensional MFCC coefficients from the microphone; output is a 2-dimensional vector [K B ], representing stiffness (N / m) and damping (Ns / m) respectively; the training results achieved a prediction accuracy of 95% on the test set.
[0072] Phase 2, Real-time Polishing
[0073] The robot performs grinding from the flat area to the raised area.
[0074] At time t0, during stable grinding of the planar area, the sensor reading is: visual: forward curvature C = 0.005 mm. -1 Normal vector gradient G = 0.02 rad / mm; Accelerometer: vibration RMS = 0.15g, frequency band energy = 0.008 g 2 / Hz, microphone: MFCC features show a stable polished spectrum. After inputting the feature sequence, the LSTM decision output parameters K =6000 N / m, B =120Ns / m, the control effect is to maintain excellent force tracking performance in the planar area, and the actual contact force is 50.2±0.5N.
[0075] At time t1, a boss transition was detected, with 60ms remaining until contact. A visual warning was issued, indicating a sharp change in curvature to C=0.08 mm. -1 The normal vector gradient G = 0.35 rad / mm; feature changes, with a significant increase in curvature and gradient components in the eigenvectors; LSTM prediction, based on the upcoming geometric mutation, outputs a new optimal parameter K. =3200 N / m, B =280 Ns / m; the system completes a smooth parameter transition 40ms before contacting the boss.
[0076] At time t2, grinding of the boss area begins. With traditional fixed parameters, the peak impact force reaches 68N due to the stiffness being maintained at 6000 N / m, and then stabilizes after 5-6 oscillations. The method of this invention, having switched to a "low stiffness, high damping" mode, has a maximum impact force of only 52.5N, and immediately stabilizes at 50.5±0.8N after two slight oscillations.
[0077] Phase 3, Online Adaptive Optimization Example
[0078] The diameter of the grinding tool (grinding wheel) decreased from 100mm to 92mm due to wear. In the early stage of wear, the LSTM prediction parameters showed a slight deviation, and the force control accuracy dropped to RMSE=1.5N. The system automatically recorded 100 sets of grinding data (features + performance) under the new working conditions. After a round of incremental learning, the model readjusted to the tool wear state, and the force control accuracy recovered to RMSE=1.0N.
[0079] This embodiment fully demonstrates the significant advantages of the present invention in constant force grinding of complex curved surfaces, realizing a technological leap from "experience-dependent" to "data-driven".
[0080] The innovation of this invention lies in:
[0081] The fusion mechanism of visual foresight and physical perception: using 3D cameras to acquire geometric information in advance, providing "predictability" for parameter adjustment.
[0082] For dedicated feature engineering for force control parameter tuning: clearly select features that are strongly correlated with force control dynamic performance, such as curvature, vibration RMS, and MFCC.
[0083] Parameter mapping based on temporal deep learning model (LSTM): Using LSTM to handle temporal dependencies and achieve dynamic tuning.
Claims
1. A method for self-tuning dynamic parameters of constant force grinding, characterized in that, Includes the following steps: 1) When the robot moves along the predetermined trajectory, multimodal sensor data is collected in real time, and data features are extracted; 2) Using the extracted multimodal feature vectors as input, the parameters of the underlying robot force controller are predicted through a pre-trained deep learning model; 3) The robot force controller uses new parameters and contact force feedback signals to drive the industrial robot to perform grinding operations through impedance control algorithms and generate joint torque commands.
2. The method for self-tuning dynamic parameters of constant force grinding according to claim 1, characterized in that, The multimodal sensing data includes: 3D point cloud data of the workpiece surface in front of the robot's motion path, torque data between the robot's end flange and the constant force actuator, three-axis acceleration data of the actuator body, and grinding audio data.
3. The method for self-tuning dynamic parameters of constant force grinding according to claim 1, characterized in that, The extracted data features are specifically as follows: Visual feature extraction: For point cloud data, calculate the average curvature C and the gradient of the normal vector change of the path segment in front of the center point of the robot tool with a set distance and length; Physical feature extraction: Calculate the RMS value A of the acceleration signal over a set time period in the past. rms and the frequency band energy A at the set frequency freq And calculate the Mel-frequency cepstral coefficient characteristics M1-M13 of the audio signal within the same time period; feature Fusion: Combining C, G, and A rms A freq The 17 eigenvalues M1-M13 are concatenated to form an eigenvector V. t .
4. The method for self-tuning dynamic parameters of constant force grinding according to claim 1, characterized in that, Step 2) specifically refers to: The feature vector V of the current time and the past 9 time points. t V t-1 , ..., V t-9 A 10×17 dimensional matrix is constructed in chronological order and used as the input to an LSTM network. After training, the LSTM network outputs the optimal stiffness parameter K. and damping parameter B .
5. The method for self-tuning dynamic parameters of constant force grinding according to claim 1, characterized in that, Step 3) specifically refers to: The robot force controller performs low-pass filtering, zero-point calibration, and tool gravity compensation on the actual contact force collected by the torque sensor and the position and speed feedback signals of the joint encoder to obtain the real workpiece contact force. The actual pose of the robot's end effector is solved by forward kinematics, and the coordinate system reference of all computational quantities is unified. By substituting the deviation between the desired trajectory and the actual pose, as well as the contact force deviation, into the Cartesian space impedance control model through impedance control calculation, and combining the current stiffness and damping parameters, the required correction force of the robot end effector is calculated. When the actual contact force is greater than the target value, the model generates a corrective force in the yielding direction, driving the end effector to retreat accordingly; When the contact force is less than the target value, a corrective force is generated in the feed direction to drive end pressure compensation; The robot force controller synchronously calculates the feedforward torque through the inverse dynamics model, and then converts the modified force in Cartesian space into the compensation torque of each joint through the force mapping relationship of the Jacobian matrix. The torque is then superimposed with the feedforward torque to generate the final joint torque command, which is sent to the joint servo driver for execution via the real-time bus.
6. The method for self-tuning dynamic parameters of constant force grinding according to claim 1, characterized in that, It also includes an online adaptive optimization step, specifically: Record the force control performance index after each force control parameter adjustment, periodically use a batch of new sensor data sequences and their corresponding force control performance indexes as incremental training samples to incrementally learn the deep learning model, update the model weights, load the updated model weights into the robot force controller during the robot's idle period, and switch to the new model before the next workpiece grinding begins.
7. A constant force grinding dynamic parameter self-tuning system, characterized in that, include: The feature acquisition and extraction module is used to acquire multimodal sensor data in real time and extract data features when the robot moves along a predetermined trajectory; The deep learning decision module is used to predict the parameters of the underlying robot force controller by taking the extracted multimodal feature vectors as input and using a pre-trained deep learning model. The robot force controller module is used by the robot force controller to use new parameters and contact force feedback signals, through impedance control algorithms, to generate joint torque commands and drive the industrial robot to perform grinding operations.
8. A constant force grinding dynamic parameter self-tuning system according to claim 7, characterized in that, It also includes a reinforcement learning module, which records the force control performance index after each force control parameter adjustment, periodically uses a batch of new sensor data sequences and their corresponding force control performance indexes as incremental training samples to incrementally learn the deep learning model, update the model weights, load the updated model weights into the robot force controller during the robot's idle period, and switch to the new model before the next workpiece grinding begins.