Joint robot gap error compensation method based on physical information network

Abstract

A kind of joint robot gap error compensation method based on physical information network, the dynamics residual error of robot is obtained by collecting the angle, angular velocity and current data of each joint of robot in the process of robot excitation experiment and calculating, and data set is generated, for training the error detection and classification network constructed;Again construct the gap error model based on Gaussian base function and establish the gap error compensation network based on PINN, after training by data set, the trained gap error compensation framework based on PINN is used in online stage to predict error amplitude and carry out error compensation.The present application can avoid modeling the complex force contact process of joint gap, by learning the potential mapping relationship between the joint motion state of robot and gap error, the effective compensation of high-performance joint dynamics gap error is realized, and the precision and robustness of robot control technology based on dynamics model are improved.

Description

Technical Field

[0001] This invention relates to a technology in the field of robot control, specifically a method for compensating backlash error in multi-joint robots based on physical information networks. Background Technology

[0002] Accurate joint dynamics models and torque feedback are effective means to improve robot control performance. However, nonlinear factors such as joint gear backlash can affect the accuracy of robot joint servo tracking and dynamics models, thereby reducing the robot's actual control performance and affecting the implementation of specific functions such as collision detection. This is especially true for multi-joint serial robots, where the serial motion structure further amplifies the negative impact of joint backlash. Existing industrial robot dynamics and friction reconstruction technologies, because they do not consider the impact of joint backlash on robot motion and force performance, cannot effectively detect the occurrence of robot joint backlash errors, making it difficult to effectively model and compensate for the dynamic errors caused by robot joint backlash. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a method for compensating gap errors in multi-joint robots based on physical information networks. This method avoids modeling the complex force contact process of joint gaps and achieves effective compensation for high-performance joint dynamic gap errors by learning the potential mapping relationship between the robot's joint motion state and gap errors. This improves the accuracy and robustness of robot control technology based on dynamic models.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a gap error compensation method for multi-joint robots based on physical information networks. It involves collecting data on the robot's joint angles, angular velocities, and currents during an excitation experiment, calculating the robot's dynamic residual error, and generating a dataset for training an error detection and classification network. A gap error model based on the Gaussian function is then constructed, and a PINN-based gap error compensation network is established. After training with the dataset, the trained PINN-based gap error compensation framework is used in the online phase to predict the error amplitude and perform error compensation.

[0006] This invention relates to a system for implementing the above-mentioned method, comprising: a dataset construction unit, a model training unit, and an online compensation unit. The dataset construction unit collects joint motion data and current or torque data through robot excitation experiments, models and identifies robot dynamics and friction, and calculates the robot's dynamic residual error. Then, by analyzing the relationship between joint clearance and joint motion, clearance error is separated from the dynamic error, and the dynamic clearance error is classified and labeled according to joint motion state criteria and repeatability criteria, ultimately generating a dataset for online detection and classification of robot dynamic clearance errors. The model training unit performs time-series detection and classification of dynamic clearance errors by learning the historical information of robot joint motion. Simultaneously, a robot dynamic clearance error model is constructed based on the Gaussian function, and corresponding PINN compensation network branches are designed for different clearance error types and trained until convergence, resulting in a PINN compensation network model for robot dynamic clearance errors. The online compensation unit deploys the trained PINN compensation network model to the robot control host computer to realize online inference and compensation of robot dynamic clearance errors. Technical effect

[0007] This invention separates clearance error from dynamic errors based on the correlation between robot joint clearance and joint motion. Through a dynamic clearance error model based on the Gaussian function, the parameters of the error model are used as training targets for a neural network to construct a corresponding PINN framework, achieving precise compensation for dynamic clearance errors. Compared to existing methods that do not consider the potential mapping relationship between robot joint motion states and clearance errors, this invention enables real-time detection and classification of robot dynamic clearance errors, as well as accurate prediction of dynamic clearance error signals, achieving precise compensation for dynamic clearance errors. Attached Figure Description

[0008] Figure 1 This is a flowchart of the present invention;

[0009] Figure 2 This is a hardware architecture diagram for an example embodiment;

[0010] Figures 3 to 8 The following is a diagram showing the results of gap error detection and classification in an example.

[0011] Figure 9 This is a schematic diagram of a PINN-based gap error compensation network for an example.

[0012] Figure 10 The diagram shows the results of gap error compensation in the example. Detailed Implementation

[0013] like Figure 1As shown in this embodiment, a method for compensating for gap errors in a multi-joint robot based on a physical information network is involved.

[0014] Step one involves conducting robot stimulation experiments and collecting joint data, specifically including:

[0015] 1.1) The excitation trajectory of the robot's motion is generated using a fifth-order Fourier series. The motion trajectory of the i-th joint is specifically as follows: Where: a i,k and b i,k Let q be the coefficient of the k-th trigonometric function of the i-th joint. i,0 To ensure that the trajectory satisfies the constant terms of the initial position and velocity constraints, ω f Let t be the fundamental frequency of the excitation trajectory, and t be the time.

[0016] The excitation trajectory in the form of the fifth-order Fourier series is multi-differentiable, 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.

[0017] 1.2) Using the Patternsearch toolbox in MATLAB, the excitation trajectory of the robot motion is further optimized by the constraint condition number method to cover more motion states. Specifically, under the constraints of the robot joint's extreme position, velocity, acceleration and initial velocity of zero, the excitation trajectory with the minimum observation matrix condition number is obtained by the optimization algorithm in Patternsearch.

[0018] 1.3) Conduct robot excitation experiments based on the optimized excitation trajectory. Use the robot joint end encoder to collect the angle and angular velocity information of each joint. Use the robot joint actuator to obtain the drive current signal of each joint. Calculate the estimated torque of each joint based on the drive current signal, specifically: τ = krI, where: I is the drive current of each joint, k is the motor torque constant of each joint, r is the transmission proportional coefficient of each joint, and τ is the estimated torque of each joint. The values ​​of k and r are provided by the robot manufacturer or obtained independently.

[0019] 1.4) The collected angle and angular velocity information of each joint of the robot and the current estimated torque are sequentially subjected to mean filtering, low-pass filtering and differential processing to obtain the angular acceleration of each joint.

[0020] The low-pass filter is preferably a fifth-order Butterworth low-pass filter.

[0021] Step two involves performing robot dynamics and friction modeling and calculating joint torque residuals, specifically including:

[0022] 2.1) Perform robot dynamics modeling, specifically: robot joint dynamic torques. Where: q, These represent the robot's joint angles, angular velocities, and angular accelerations, respectively, and M(q) is the robot's mass matrix. The forces generated by the centripetal force and the Coriolis force, G(q) is gravity, and τ is the force generated by the centripetal force and the Coriolis force. dyn These are the dynamic torques of each joint of the robot.

[0023] 2.2) Perform friction modeling for the robot, specifically: the frictional force at the k-th joint of the robot. Where: friction model matrix α k These are empirical parameters, where sign(·) is the sign function, and the friction coefficient is the coefficient of friction. These are the Coulomb friction coefficient, viscous friction coefficient, and friction offset of the k-th joint, respectively.

[0024] 2.3) Solve the constructed robot dynamics model and friction model using a least-squares-based parameter identification method or a PINN-based model training method to obtain the dynamic torque τ of each joint of the robot. dyn and frictional torque τ fri Then, calculate the torque residual τ of each joint of the robot. err =τ-τ dyn -τ fri .

[0025] Step 3: Construct a robot dynamics backlash error classification dataset based on the robot's motion state, specifically including:

[0026] 3.1) Generate feature factors based on the robot's motion state, specifically: calculate the feature factors representing the motion state of the robot's k-th joint. Where: t i and These represent the time and joint velocity corresponding to the i-th motion state sample, respectively; when Then it is the tth i At a certain moment, the direction of motion of the robot's k-th joint reverses.

[0027] 3.2) Generate backlash error tags based on the robot's state. Wherein: κ i =1 indicates that a commutation error exists at this moment, when κ i =2 indicates that there is a start / stop error at this moment, when κ i =0 indicates that there is no error caused by the gap at that moment, and the duplicate labels corresponding to the same gap error sample are further screened out according to the repeatability criterion.

[0028] When robot dynamic backlash error occurs, its proportion is much larger than other error components in the torque residual. Therefore, the robot's dynamic backlash error can be approximated as: κ i =1 or κ i A value of 2 indicates the existence of dynamic clearance error, which can be approximated as torque residual, i.e., the torque residual τ of each joint of the robot calculated in step 2.3. err .

[0029] 3.3) Constructing a robot dynamics backlash error classification dataset: This dataset includes the joint angles q and angular velocities preprocessed in step one. and angular acceleration Calculate the robot backlash error label κ for each time step. Use a sliding window of length 32 to calculate q. A sliding window approach is used, where the joint angles, angular velocities, and angular accelerations of each robot joint at the current moment are used as samples in the dataset; the gap error label corresponding to the current moment is used as the sample label. For the dataset collected over a period of time, it is divided into training, validation, and test sets in proportions of 60%, 20%, and 20%, respectively.

[0030] Step four: Based on the dataset generated in step three, a temporal convolutional network (TCN) for classification is trained to achieve online detection and category prediction of gap errors. The joint angles, angular velocities, and angular accelerations of each joint of the robot at the current time are used as inputs, and the gap error label corresponding to the current time is used as the prediction result. The historical information of the robot's motion state is extracted by stacked one-dimensional convolutional architecture and dilated convolution to detect whether an error has occurred and to determine the category of the error.

[0031] The dilated convolution in:* d For the extended convolution operation, x is the input data, f is the extended convolution filter, k is the filter size, and d is the dilation factor.

[0032] The prediction results in: This is the predicted label value for the gap error category. When this predicted label value changes from 0 to a non-zero value at a certain moment, it indicates that a gap error will occur at that moment. Real-time online detection of joint gap errors is achieved by predicting the gap error category. Figures 3-8 The figure shows the detection and classification results of the gap error of each joint of the robot.

[0033] Step 5: Construct a gap error model based on the Gaussian function and establish a gap error compensation network based on PINN to predict the error magnitude and perform error compensation. Specifically, this includes:

[0034] 5.1) The j-th gap error sample Φ j (S j ,δ Rev )=Φ j (t), t∈B(S) j ,δ Rev ),in: B(S j ,δ Rev ) represents the time interval of the gap error, i.e., the j-th consecutive interval where the error category label κ is not 0, a j b j c j These parameters determine the shape of the Gaussian function, and are updated through learning to ensure that the Gaussian function can accurately fit the gap error. The length of a single gap error sample in the time domain is finite.

[0035] 5.2) The mapping relationship between the robot's motion state and the Gaussian function morphological parameters of the gap error is learned through a multilayer perceptron (MLP), and the gap errors in steps 2-4 are classified to obtain the predicted value of the gap error during a certain motion period: Specifically: Predicted value in: These are predicted values ​​for the parameters of the Gaussian function, including the joint angle q and angular acceleration. As input to the model, joint angles significantly affect the temporal offset of the gap error, while angular acceleration has a greater impact on the peak value of the gap error. κ represents the error type; different branches of the MLP are used to learn different types of errors to model the two types of gap errors with different characteristics separately. θ represents the network parameters, which are updated during model training. and These are the predictions for the j-th commutation error sample and the start / stop error sample, respectively.

[0036] Through specific experiments, in such Figure 2 With the hardware setup shown—a six-axis serial robot body, robot control cabinet, host computer with TwinCAT3 software, Beckhoff controller, and EtherCAT communication bus—the robot dynamics gap error model based on the Gaussian function described in step 5.1 is constructed in the PyTorch environment, and the corresponding PINN gap error compensation network is built. Since gap errors typically have higher peak values ​​and sparser time-domain distributions compared to dynamic errors caused by other factors, the loss function of the network is designed as follows: in: The weighting coefficients of the loss function emphasize the impact of larger errors, thus achieving a better approximation of the gap error. σ is an integer close to 0, introduced to prevent computational errors when the error is zero. Adam is selected as the optimizer, with a learning rate of 0.01, and the PINN gap error compensation network model is trained until convergence. This invention predicts the parameters of the gap error model through the PINN network, rather than directly predicting the gap error value. This effectively simplifies the complexity of the mapping relationship, making the proposed error compensation method more robust and easier to learn.

[0037] A new robot motion trajectory is selected, and the gap error online detection and classification method in step three is used for real-time online detection. When a gap error is detected, the trained PINN gap error compensation network model is used for online inference and compensation of the gap error.

[0038] Through specific practical experiments, under the aforementioned hardware environment settings, the root mean square error (RMSE) values ​​of the torque prediction errors for each joint of the robot decreased from 2.2218 N·m, 1.8042 N·m, 1.0625 N·m, 0.5756 N·m, 0.2578 N·m, and 0.2551 N·m before compensation to 1.4213 N·m, 1.4343 N·m, 0.6292 N·m, 0.2745 N·m, 0.1356 N·m, and 0.6775 N·m. Comparisons were also made with other existing methods, including MLP-based methods, TCN-based methods, Convolutional Neural Network (CNN)-based methods, and Long Short-Term Memory (LSTM)-based methods, such as... Figure 10 As shown, the results indicate that the present invention can achieve high-precision gap error compensation, which is superior to existing methods. The comparison results are shown in Table 1.

[0039] Table 1. Comparison of RMSE (N·m) of gap error compensation effects

[0040] Compared with existing technologies, this invention utilizes a robot dynamics clearance error model based on Gaussian functions. This model effectively fits the actual robot dynamics clearance error, improving the compensation effect of the clearance error and achieving higher accuracy in predicting joint dynamic torques. Instead of directly predicting the clearance error value, the parameters of the clearance error model are predicted using a PINN network, which effectively simplifies the complexity of the mapping relationship, making the proposed error compensation method more robust and easier to learn, thus ensuring the overall motion control performance of the industrial robot based on the dynamics model.

[0041] 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 compensating backlash error in a multi-joint robot based on a physical information network, characterized in that, By collecting data on the robot's joint angles, angular velocities, and currents during the robot excitation experiment, and calculating the robot's dynamic residual error, a dataset is generated to train the constructed error detection and classification network. A gap error model based on Gaussian function is constructed and a gap error compensation network based on PINN is established. After training with the dataset, the trained gap error compensation framework based on PINN is used in the online stage to predict the error magnitude and perform error compensation. The robot stimulation experiment described above is achieved in the following way: 1.1) The excitation trajectory of the robot motion is generated using a fifth-order Fourier series. The specific trajectory of joint movement is as follows: ,in: and For the first The first joint The coefficients of the trigonometric functions, To ensure that the trajectory satisfies the constant terms of the initial position and velocity constraints, The fundamental frequency of the excitation trajectory, For time; 1.2) Using the Patternsearch toolbox in MATLAB, the excitation trajectory of the robot motion is further optimized by the constraint condition number method to cover more motion states. Specifically, under the constraints of the robot joint's extreme position, velocity, acceleration and initial velocity of zero, the excitation trajectory with the minimum observation matrix condition number is obtained by the optimization algorithm in Patternsearch. 1.3) Conduct robot excitation experiments based on the optimized excitation trajectory. Use the robot joint end encoder to collect the angle and angular velocity information of each joint, and use the robot joint actuator to obtain the drive current signal of each joint. Calculate the torque of each joint based on the drive current signal, specifically: ,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; 1.4) The collected angle and angular velocity information of each joint of the robot and the current estimated torque are sequentially subjected to mean filtering, low-pass filtering and differential processing to obtain the angular acceleration of each joint.

2. The method for compensating backlash error in a multi-joint robot based on a physical information network according to claim 1, characterized in that, The aforementioned dynamic residual error is calculated in the following manner: 2.1) Perform robot dynamics modeling, specifically: robot joint dynamic torques. ,in: , , These are the robot's joint angles, angular velocities, and angular accelerations, respectively. The mass matrix of the robot, The forces generated by centripetal force and Coriolis force, For gravity, These are the dynamic torques of the robot's joints; 2.2) Perform robot friction modeling, specifically: the robot's first... Friction of joints Where: friction model matrix , , , , These are empirical parameters, The sign function is the coefficient of friction. , , , The first Coulomb friction coefficient, viscous friction coefficient, and frictional bias of the joint; 2.3) Solve the constructed robot dynamics model and friction model using a least-squares-based parameter identification method or a PINN-based model training method to obtain the dynamic torques of each joint of the robot. and frictional torque Then, the torque residuals of each joint of the robot are calculated. .

3. The method for compensating backlash error in a multi-joint robot based on a physical information network according to claim 1, characterized in that, The dataset was obtained in the following way: 3.1) Generate feature factors based on the robot's motion state, specifically: calculate the feature factors representing the robot's motion state. Characteristic factors of joint motion state ,in: and The first The time and joint velocity corresponding to each motion state sample; when Then it is the first Time Robot The direction of joint movement has reversed; 3.2) Generate backlash error tags based on the robot's state. ,in: This indicates that a commutation error exists at that moment. This indicates that there is a start / stop error at that moment. This indicates that there is no error caused by the gap at that moment, and further, duplicate labels corresponding to the same gap error sample are screened out according to the repeatability criterion; When robot dynamic backlash error occurs, its proportion is much larger than other error components in the torque residual. Therefore, the robot's dynamic backlash error can be approximated as follows: or This indicates the existence of dynamic clearance error, which can be approximated as torque residual, i.e., the torque residual of each joint of the robot calculated in step 2.

3. ; 3.3) Constructing a robot dynamics backlash error classification dataset: This dataset is generated from the joint angles preprocessed in step one. angular velocity and angular acceleration Calculate the robot backlash error label for each time step. Using a sliding window of length 32 to , , A sliding window process is used, which means that the joint angles, angular velocities, and angular accelerations of each joint of the robot at the current time are used as samples of the dataset; the gap error label corresponding to the current time is used as the sample label. For the dataset collected within a certain period of time, it is divided into training set, validation set and test set according to the proportions of 60%, 20% and 20%.

4. The method for compensating backlash error in a multi-joint robot based on a physical information network according to claim 1, characterized in that, The error detection and classification network constructed by the training refers to: training a temporal convolutional network (TCN) for classification based on a dataset to achieve online detection and category prediction of gap errors: taking the joint angles, angular velocities, and angular accelerations of each joint of the robot at the current time 32 times before the current time as input, and taking the gap error label corresponding to the current time as the prediction result, extracting the historical information of the robot's motion state through a stacked one-dimensional convolutional architecture and dilated convolution, detecting whether an error has occurred and determining the category of the error.

5. The method for compensating backlash error in a multi-joint robot based on a physical information network according to claim 4, characterized in that, The dilated convolution ,in: To expand the convolution operation, For input data, For dilated convolutional filters, It is the filter size. It is an expansion factor; The prediction results ,in: This is the predicted label value for the gap error category. When the predicted label value changes from 0 to non-zero at a certain moment, it indicates that a gap error will occur at that moment. Real-time online detection of joint gap errors is achieved by predicting the gap error category.

6. The method for compensating backlash error in a multi-joint robot based on a physical information network according to claim 1, characterized in that, The construction of the gap error model based on the Gaussian function and the establishment of the gap error compensation network based on PINN specifically include: 5.1) Section One gap error sample ,in: , The time interval during which the gap error persists, i.e., the error category label. The non-zero first A continuous interval, , , These are the parameters that determine the shape of the Gaussian function. By learning and updating them, we can ensure that the Gaussian function can accurately fit the gap error. 5.2) The mapping relationship between the robot's motion state and the Gaussian morphological parameters of the gap error is learned through a multilayer perceptron (MLP), and the gap error is classified and predicted: Specifically: Predicted value ,in: , , These are the predicted values ​​of the Gaussian function parameters, and the joint angles. and angular acceleration As input to the model, joint angles have a significant impact on the time-domain shift of clearance error, while angular acceleration has a significant impact on the peak value of clearance error. To address the different types of errors, different branches of the MLP are used for learning, enabling separate modeling of two types of gap errors with different characteristics. These are network parameters that are updated during model training. and The first Prediction of commutation error samples and start-stop error samples.

7. A multi-joint robot gap error compensation system based on a physical information network for implementing the method of any one of claims 1-6, characterized in that, include: The system comprises a dataset construction unit, a model training unit, and an online compensation unit. The dataset construction unit collects joint motion data and current or torque data through robot excitation experiments, models and identifies robot dynamics and friction, calculates the robot's dynamic residual error, and then separates the clearance error from the dynamic error by analyzing the relationship between joint clearance and joint motion. Based on joint motion state criteria and repeatability criteria, the dynamic clearance error is classified and labeled, ultimately generating a dataset for online detection and classification of robot dynamic clearance errors. The model training unit performs temporal detection and classification of dynamic clearance errors by learning the historical information of robot joint motion. Simultaneously, it constructs a robot dynamic clearance error model based on the Gaussian function and designs corresponding PINN compensation network branches for different clearance error types, training them until convergence to obtain the PINN compensation network model for robot dynamic clearance errors. The online compensation unit deploys the trained PINN compensation network model onto the robot control host computer to realize online inference and compensation of robot dynamics backlash errors.

Images