Industrial robot dynamics and friction reconstruction method based on pinn
By combining physical information neural networks and Lagrange dynamics models, the problem of handling nonlinear friction in industrial robot dynamics is solved, achieving high-precision joint torque prediction and improving motion control performance.
Patent Information
- Application Number
- CN202411455660.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-18
- Publication Date
- 2025-11-18
- Estimated Expiration
- 2044-10-18
AI Technical Summary
Existing industrial robot dynamics models struggle to linearize nonlinear components such as joint friction, leading to cumbersome modeling and difficulty in achieving high-precision joint torque prediction, thus impacting motion control performance.
A physical information neural network (PINN) is used in conjunction with Lagrange dynamics and Coulomb viscous friction model. A hybrid learning strategy is used to construct robot dynamics and friction models, which simplifies the modeling process and achieves high-precision joint torque prediction.
It simplifies the modeling and prediction process of robot joint torques, improves the overall motion control performance, and achieves high-precision joint torque prediction.
Smart Images
Figure CN119335901B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a technology in the field of automated manufacturing, specifically an industrial robot dynamics and friction reconstruction method based on Physical Information Neural Network (PINN). Background Technology
[0002] Currently, the dynamics models of industrial robots are widely used in specific scenarios such as robot controller design, motion planning with joint force / torque constraints, offline simulation, drag teaching, and the design of disturbance observers. Existing modeling techniques rely on cumbersome formula derivations and strict optimization constraints, and have difficulty handling the highly nonlinear components of joint torques that are difficult to linearize, such as nonlinear joint friction. Summary of the Invention
[0003] This invention addresses the shortcomings of existing technologies in modeling industrial robot dynamics for specific task scenarios and in optimizing the dynamic and frictional components of industrial robot dynamics. It proposes a physical information neural network-based method for industrial robot dynamics modeling. This method considers a hybrid learning strategy for the dynamic and frictional components of robot dynamics. By constructing a physical information neural network, it combines the Lagrangian dynamics modeling method with deep learning methods. While leveraging the powerful learning capabilities of neural networks for nonlinear features, it strictly ensures the physical constraints of robot joint dynamics. Furthermore, it simultaneously considers the modeling of nonlinear friction in industrial robot joints. Through the implementation of this invention, the cumbersome derivation process of traditional robot dynamics modeling methods can be avoided, significantly simplifying the modeling and prediction process of joint torques in industrial robots. It also achieves high-precision joint torque prediction, thereby improving the overall motion control performance of industrial robots based on dynamic models.
[0004] This invention is achieved through the following technical solution:
[0005] This invention relates to a method for modeling the dynamics of industrial robots based on physical information neural networks. In the offline stage, the robot excitation trajectory is constructed and the robot excitation experiment is set up. At the same time, robot joint position data, velocity data and current data are collected during the process to construct a dataset for hybrid learning of the PINN model for robot dynamics modeling based on the Lagrange method and the PINN model for robot friction modeling based on the Coulomb viscosity model. In the online stage, the trained robot dynamics PINN model and robot friction PINN model are used for real-time inference and prediction to obtain the dynamic components and frictional components of the robot joint torque.
[0006] This invention relates to a system for implementing the above-mentioned method, comprising: a model building unit, a hybrid learning unit, and an online inference unit, wherein: the model building unit establishes a robot dynamics model based on robot Lagrange dynamics, establishes a robot joint friction model based on Coulomb viscous friction model, and constructs a physical information neural network of robot dynamics and friction based on the model to obtain the corresponding network model; the hybrid learning unit, based on the robot dynamics model and friction model, designs an iterative hybrid learning strategy for the difficult-to-distinguish dynamic and frictional components in robot joint torques, achieving high-precision learning and prediction of the two torque components; the online inference unit deploys the trained PINN network model to the robot control host computer to realize online inference and prediction of robot joint torques.
[0007] Technical effect
[0008] This invention constructs a physical information neural network for robot dynamics modeling based on the Lagrange method. It replaces the partial differential equation process of the Lagrange method with a deep network layer based on chain-like computation, and considers the physical feasibility constraints of the robot dynamics model in the design of the physical information neural network. Then, within the unified physical information neural network framework, it introduces modeling and prediction of nonlinear friction of robot joints. Finally, through a cyclical iterative hybrid learning strategy, it simultaneously learns the dynamic and frictional components of robot joint torques, achieving overall dynamic modeling of industrial robot joint torques. Compared with existing technologies, this invention ensures the physical constraints of robot joint dynamics, avoids the cumbersome derivation process of traditional robot dynamics modeling methods, greatly simplifies the modeling and prediction process of industrial robot joint torques, and achieves high-precision joint torque prediction, thereby improving the overall motion control performance of industrial robots based on dynamic models. Attached Figure Description
[0009] Figure 1 This is a flowchart of the present invention;
[0010] Figure 2 This is a hardware framework diagram for an embodiment;
[0011] Figure 3 This is a network schematic diagram for an example.
[0012] Figure 4 Here is a flowchart of the algorithm for an example implementation;
[0013] Figures 5 to 10 This is a model verification diagram for an example. Detailed Implementation
[0014] like Figure 1 As shown in the figure, this embodiment relates to a dynamic modeling method for industrial robots based on physical information neural networks, including:
[0015] Step one, constructing and optimizing the excitation trajectory of the robot's motion, specifically includes:
[0016] 1.1) The excitation trajectory for robot motion is generated using a fifth-order Fourier series, specifically: ,in: The order of the selected Fourier series , and For the first The first joint The coefficients of the trigonometric functions, These are constants used to ensure that the trajectory meets the initial position and velocity constraints. The fundamental frequency of the excitation trajectory. For time.
[0017] The excitation trajectory in the form of a fifth-order Fourier series is differentiable in many orders, making it easy to obtain analytical expressions for velocity and acceleration. Furthermore, the flexibility effect of the robot can be avoided by designing the frequency range.
[0018] 1.2) Using the Patternsearch toolbox in MATLAB, the robot's motion trajectory is further optimized by the constraint condition number method to cover more motion states. Specifically, under the constraints of the robot's joint limit position, velocity, acceleration, and initial velocity of zero, the optimization algorithm in Patternsearch is used to find the excitation trajectory with the minimum observation matrix condition number. This indicates that the trajectory is more general and covers more motion states.
[0019] Step two involves conducting a robot excitation experiment, collecting data on robot joint position, velocity, and current, and estimating the torque based on the current at each joint. This includes:
[0020] 2.1) Settings Figure 2 The robot executes the excitation trajectory generated in step one, uses the robot joint end encoder to collect the position and velocity information of each joint of the robot, and uses the robot joint actuator to get the drive current signal of each joint of the robot.
[0021] 2.2) Calculate the current of each joint of the robot to estimate the torque. ,in: This refers to the driving current for each joint of the robot. Let be the motor torque constant for each joint. These are the transmission ratio coefficients for each joint. Estimate the torque for the current in each joint. and The values are provided by the robot manufacturers or identified by the manufacturers themselves.
[0022] Step 3: Filter the data collected in Step 2 and divide the dataset, specifically including:
[0023] 3.1) First, the robot joint position data, velocity data and current estimated torque collected in step two are subjected to mean filtering. Then, a fifth-order Butterworth low-pass filter is used to further filter the robot joint velocity data and perform differential processing to obtain the acceleration information of each joint.
[0024] 3.2) The robot joint positions, velocities, accelerations and current estimated torques after two filtering steps are used as the dataset for robot dynamics modeling and PINN network training, and are divided into training set, validation set and test set according to a 6:2:2 ratio.
[0025] Step four, construct the following based on the Lagrange method: Figure 3 The PINN model shown for robot dynamics modeling specifically includes:
[0026] 4.1) The second-order ordinary differential equations of the robot's rigid body dynamics model can be derived using the Lagrange method. ,in: These are the robot's joint angles, velocities, and accelerations, respectively. The mass matrix of the robot, It is a force generated by centripetal force and Coriolis force. For gravity, These are the dynamic torques of each joint of the robot.
[0027] 4.2) Using PINN to learn the various parts of the dynamics model: the robot's mass matrix Decomposed into ,in: It is a lower triangular matrix, used for network computation. and The result is , ,in: express The network prediction value, For the corresponding network parameters, express The network prediction value, For the corresponding network parameters, express The network prediction values; the optimization problem for training the PINN network in the dynamics part is: ,in: The constraints are , This represents the inverse dynamics model of a robot, which uses joint angles, velocities, and accelerations. and network parameters Calculate the predicted dynamic torque value , Let represent the loss function of the PINN network. The lower triangular matrix can be further decomposed into... ,in Indicates off-diagonal elements. Represents diagonal elements, in order to ensure The diagonal and Positive eigenvalues, The network output layer uses a non-negative activation layer (such as ReLU), while The network output layer can be a linear activation layer.
[0028] 4.3) For robot dynamics models that require solving partial differential equations The term is obtained by chaining computation using a PINN network: ,in: It can be by The results were obtained. Therefore, the robot inverse dynamics model All derivatives contained therein can be calculated in closed-form, thus the dynamic torque of the robot can be obtained. .
[0029] Step 5: Construct the PINN model for robot friction modeling based on the Coulomb viscosity model, specifically as follows: ,in: The friction component representing the joint torque of a robot. , and They represent the robot's number 1 and 2. The Coulomb coefficient of friction, the viscous friction coefficient, and the frictional bias constant of the joint. Represents a symbolic function, indicating Zero matrix. Represents the friction coefficient matrix. This represents the friction model matrix.
[0030] Step six, as follows Figure 4 The hybrid learning strategy for the robot dynamics model and friction model shown includes:
[0031] 6.1 Simultaneously train the dynamics and friction components to obtain the dynamic torque. and predicted friction torque ,in: and These represent the prediction results from the dynamic model and the friction model, respectively. and The PINN network models represent the robot's dynamics and friction components, respectively; the network's loss function. .
[0032] 6.2 Error backpropagation is performed, and the network parameters of the dynamics and friction models are updated. This process is repeated multiple times until... The function converges.
[0033] 6.3 Freeze the network parameters for the friction component and train only the dynamics component to obtain the dynamic torque. and predicted friction torque ,in: The PINN network friction model represents a frozen robot, where the network parameters are no longer updated; at this point, the network's loss function... .
[0034] 6.4 Error backpropagation is performed, and the network parameters of the dynamic model are updated. This process is repeated multiple times until... The function converges.
[0035] 6.5 Alternately execute steps 6.1-6.4 to form the inner loop of the hybrid learning strategy for the robot's dynamics model and friction model, and train it. generation.
[0036] 6.6 After the inner loop training is complete, freeze the friction parameters and train only the dynamics part, using the same training method as the inner loop, until... Once the function converges, the training of the entire PINN network is complete.
[0037] Step 7: Train the PINN model for hybrid modeling of robot dynamics and friction. Specifically, based on the dataset obtained in Step 3, the network input consists of joint angles. Joint speed Joint acceleration And joint current estimation torque The network output is the prediction results based on the PINN network-based dynamic and frictional models. and The total joint torque is obtained. The training loss function is: ,in: The number of samples in the dataset. For the robot The joint in the first Estimating torque value based on current at each time point For the robot The joint in the first PINN network predictions of dynamic torque at each time point For the robot The joint in the first The PINN network predicts the compensated frictional torque at each time point. The loss function reflects the mean of the symmetric mean absolute percentage error (SMAPE) of the torques at each joint of the robot.
[0038] Step 8: Verify the effectiveness of the above-described robot PINN modeling method on the test trajectory: Select the validation set from the dataset generated in Step 3, and use the trained PINN dynamics and friction hybrid model to predict the torques of each joint of the robot. The prediction results and residual values are as follows: Figures 5-10 As shown in the figure. The verification results show that the torque values of the robot's six joints are predicted with high accuracy, which is better than existing methods based on dynamic parameter identification and data-driven dynamic modeling methods.
[0039] like Figure 2 The diagram shows the environment setup for this embodiment, including a system control module and a robot body execution module. The system control module includes a host computer, a Beckhoff industrial PC, an EtherCAT communication bus, and a robot control cabinet. The host computer designs the robot's motion trajectory, imports it into the Beckhoff industrial PC, and then sends it to the robot control cabinet via the EtherCAT communication bus, controlling the robot body to move according to the commanded trajectory. The robot body execution module includes servo motors and the robot body itself. The robot used in this embodiment is a self-developed industrial robot with six rotary joints. The servo motors and drivers are from the Panasonic MINAS A6 series. Based on the Beckhoff industrial PC's real-time PLC core, the system can collect servo motor status information at a maximum frequency of 20kHz. During the robot body's movement, joint data is simultaneously collected, fed back to the host computer, and stored as training trajectory data for robot dynamics modeling and PINN network learning.
[0040] Through specific practical experiments, under the aforementioned hardware environment settings, with the training batch size of the network of this invention set to 50 and the learning rate set to 5e-4, and Adam used as the optimizer, the final RMSE values of the torque prediction errors for each joint of the robot were: 6.4048 N·m, 12.7397 N·m, 3.911 N·m, 1.5428 N·m, 0.8337 N·m, and 0.5016 N·m. Here, RMSE represents the root mean square error.
[0041] Compared with existing technologies, this method reduces the torque prediction error of each joint of the robot and achieves a higher accuracy in joint torque prediction, thereby improving the overall motion control performance of industrial robots based on dynamic models.
[0042] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for industrial robot dynamics and friction reconfiguration based on PINN, characterized in that, In the offline phase, robot excitation trajectories are constructed and robot excitation experiments are set up. At the same time, robot joint position data, velocity data and current data are collected during the process to construct a dataset for hybrid learning of the PINN model for robot dynamics modeling based on the Lagrange method and the PINN model for robot friction modeling based on the Coulomb viscosity model. During the online phase, the trained robot dynamics PINN model and robot friction PINN model are used for real-time inference and prediction to obtain the dynamic components and frictional components of the robot joint torque. The PINN model for robot dynamics modeling is obtained in the following way: 4.1) The second-order ordinary differential equations of the robot's rigid body dynamics model are derived using the Lagrange method. ,in: These are the robot's joint angles, velocities, and accelerations, respectively. The mass matrix of the robot, It is a force generated by centripetal force and Coriolis force. For gravity, These are the dynamic torques of the robot's joints; 4.2) Using PINN to learn the various parts of the dynamics model: the robot's mass matrix Decomposed into ,in: It is a lower triangular matrix, used for network computation. and The result is , ,in: express The network prediction value, For the corresponding network parameters, express The network prediction value, For the corresponding network parameters, express The network prediction values; the optimization problem for training the PINN network in the dynamics part is: ,in: The constraints are , This represents the inverse dynamics model of a robot, which uses joint angles, velocities, and accelerations. and network parameters Calculate the predicted dynamic torque value , The loss function of the PINN network is further decomposed into the lower triangular matrix as follows: ,in Indicates off-diagonal elements. Represents diagonal elements, in order to ensure The diagonal and Positive eigenvalues, The network output layer uses a non-negative activation layer. The network output layer uses a linear activation layer; 4.3) For robot dynamics models that require solving partial differential equations The term is obtained by chaining computation using a PINN network: ,in: Depend on The results are obtained, therefore, the robot inverse dynamics model All derivatives included are calculated in closed-form, thus yielding the robot's dynamic torque. ; The PINN model for robot friction modeling based on the Coulomb viscosity model is obtained in the following way: ,in: The friction component representing the joint torque of a robot. , and They represent the robot's number 1 and 2. The Coulomb coefficient of friction, the viscous friction coefficient, and the frictional bias constant of the joint. Represents a symbolic function, indicating Zero matrix, Represents the friction coefficient matrix. This represents the friction model matrix.
2. The PINN-based industrial robot dynamics and friction reconfiguration method according to claim 1, characterized in that, The construction and optimization of the excitation trajectory for robot motion specifically includes: 1.1) The excitation trajectory for robot motion is generated using a fifth-order Fourier series, specifically: ,in: Let i be the order of the selected Fourier series, i = 1, 2, ..., 6. and For the first The first joint The coefficients of the trigonometric functions, These are constant terms used to ensure that the trajectory meets the initial position and velocity constraints. The fundamental frequency of the excitation trajectory, For time; 1.2) Using the Patternsearch toolbox in MATLAB, the robot's motion trajectory is further optimized by the constraint condition number method to cover more motion states. Specifically, under the constraints of the robot's joint limit position, velocity, acceleration, and initial velocity of zero, the excitation trajectory with the minimum observation matrix condition number is obtained by using the optimization algorithm in Patternsearch.
3. The PINN-based industrial robot dynamics and friction reconfiguration method according to claim 1, characterized in that, The estimated torque of the current at each joint of the robot is obtained in the following way: 2.1) Set the excitation trajectory generated in step one for the robot to execute, use the robot joint end encoder to collect the position and velocity information of each joint of the robot, and use the robot joint actuator to get the drive current signal of each joint of the robot. 2.2) Calculate the current of each joint of the robot to estimate the torque. ,in: This refers to the driving current for each joint of the robot. Let be the motor torque constant for each joint. These are the transmission ratio coefficients for each joint. Estimate the torque for the current in each joint. and The values are provided by the robot manufacturers or identified by the manufacturers themselves.
4. The PINN-based industrial robot dynamics and friction reconfiguration method according to claim 1, characterized in that, The dataset was obtained in the following way: 3.1) First, the collected robot joint position data, velocity data and current estimated torque are processed by mean filtering. Then, a fifth-order Butterworth low-pass filter is used to further filter the robot joint velocity data and perform differential processing to obtain the acceleration information of each joint. 3.2) The robot joint positions, velocities, accelerations and current estimated torques after two filtering steps are used as the dataset for robot dynamics modeling and PINN network training, and are divided into training set, validation set and test set.
5. The PINN-based industrial robot dynamics and friction reconfiguration method according to claim 1, characterized in that, The hybrid learning strategy specifically includes: 6.1 Simultaneously train the dynamics and friction components to obtain the dynamic torque. and predicted friction torque ,in: and These represent the prediction results from the dynamic model and the friction model, respectively. and The PINN network models represent the robot's dynamics and friction components, respectively; the network's loss function. ; 6.2 Error backpropagation is performed, and the network parameters of the dynamics and friction models are updated, iterating multiple times until... The function converges; 6.3 Freeze the network parameters for the friction component and train only the dynamics component to obtain the dynamic torque. and predicted friction torque ,in: The PINN network friction model represents a frozen robot, where the network parameters are no longer updated; at this point, the network's loss function... ; 6.4 Error backpropagation is performed, and the network parameters of the dynamic model are updated. This process is repeated multiple times until... The function converges; 6.5 Alternately execute steps 6.1-6.4 to form the inner loop of the hybrid learning strategy for the robot's dynamics model and friction model, and train it. generation; 6.6 After the inner loop training is complete, freeze the friction parameters and train only the dynamics part, using the same training method as the inner loop, until... Once the function converges, the training of the entire PINN network is complete.
6. The PINN-based industrial robot dynamics and friction reconfiguration method according to claim 1 or 5, characterized in that, Based on the dataset required for network training, the network input is the joint angle. Joint speed Joint acceleration And joint current estimation torque The network output is the prediction results based on the PINN network-based dynamic and frictional models. and The total joint torque is obtained. The training loss function is: ,in: The number of samples in the dataset. For the robot The joint in the first Estimating torque value based on current at each time point For the robot The joint in the first PINN network predictions of dynamic torque at each time point For the robot The joint in the first The PINN network predicts the compensated friction torque at each time point. The loss function reflects the mean of the symmetric mean absolute percentage error (SMAPE) of the torques at each joint of the robot.
7. A PINN-based industrial robot dynamics and friction reconfiguration system for implementing the method of any one of claims 1-6, characterized in that, include: The system comprises a model building unit, a hybrid learning unit, and an online inference unit. The model building unit establishes a robot dynamics model based on Lagrangian dynamics and a robot joint friction model based on the Coulomb viscous friction model. It then constructs a neural network for the physical information of robot dynamics and friction based on these models, resulting in the corresponding network model. The hybrid learning unit, building upon the robot dynamics and friction models, designs an iterative hybrid learning strategy to address the difficult-to-distinguish dynamic and frictional components in the robot joint torques, achieving high-precision learning and prediction of both torque components. The online inference unit deploys the trained PINN network model onto the robot control host computer to achieve online inference and prediction of robot joint torques.