Industrial Robot Joint Torque Compensation Method Based on Spatio-Temporal Graph Convolutional Network

Abstract

An industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network. In the offline stage, an optimized motion trajectory that meets the robot joint position and speed limits is generated. The robot is set to move along the optimized motion trajectory, and the robot joint positions, speeds, and joint currents used to calculate the estimated torques of each robot joint are synchronously collected. The robot dynamics model is obtained through robot dynamics parameter identification, and then the preliminary joint prediction torques are obtained. According to the spatial relationship between robot joints and kinematic and dynamic characteristics, a joint graph adjacency matrix is constructed, thereby obtaining a sample set for training the spatio-temporal graph convolutional network. In the online stage, the torque compensation amounts of each joint at each moment are predicted using the trained spatio-temporal graph convolutional network, that is, the difference between the preliminary joint prediction torques and the joint estimated torques. This compensation torque is fed forward to the driving motors of each robot joint, thereby effectively reducing the torque tracking error of each robot joint. The present invention significantly improves the motion performance of industrial robots.

Description

Technical Field

[0001] The present invention relates to a technology in the field of robot control, specifically an industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network. Background Art

[0002] Existing torque control compensation methods are mainly model-based control methods, which rely on the accuracy of the dynamic model established in the control method. In fact, due to influencing factors such as friction, joint flexibility, and motion coupling that are difficult to accurately model, the established dynamic model will always have uncertain disturbances, resulting in torque tracking errors that are difficult to compensate. In addition, the control methods based on robot dynamics are computationally complex and have a large amount of operations, making it difficult to achieve real-time calculation and control. Summary of the Invention

[0003] Aiming at the deficiencies that the effect of the existing torque compensation technology is restricted by the accuracy of the dynamic model, the existing technology only considers the time correlation between the kinematic and dynamic data of the robot and does not consider the spatial connection relationship and physical characteristics between the joints of the industrial robot, and cannot compensate each joint simultaneously, the present invention proposes an industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network, which considers the time continuity of the information of each joint of the robot and the spatial relationship and physical characteristics between the joints. On the basis of the robot model obtained by robot dynamics parameter identification, the joint torque tracking error of the robot is further compensated, significantly improving the motion performance of the industrial robot.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to an industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network. In the offline stage, an optimized motion trajectory that meets the robot joint position and speed limits is generated, the robot is set to move according to the optimized motion trajectory, and the robot joint positions, speeds, and joint currents used to calculate the estimated torques of each robot joint are synchronously collected. The robot dynamic model is obtained through robot dynamics parameter identification, and then the preliminary joint predicted torques are obtained. Then, according to the spatial relationship and kinematic and dynamic characteristics between the robot joints, a joint graph adjacency matrix is constructed, thereby obtaining a sample set for training the spatio-temporal graph convolutional network. In the online stage, the torque compensation amounts of each joint at each moment are predicted by using the trained spatio-temporal graph convolutional network, that is, the difference between the preliminary joint predicted torques and the joint estimated torques. This compensation torque is fed forward to the drive motors of each robot joint, thereby effectively reducing the torque tracking error of each robot joint.

[0006] The optimized motion trajectory is preferably an excitation trajectory in the form of a fifth-order Fourier series as the training trajectory for robot dynamics parameter identification and torque compensation. After optimizing the robot motion trajectory through the condition number, an optimized motion trajectory that conforms to the robot joint position and speed limits is obtained through physical constraints.

[0007] The acquisition mentioned above means that during the movement of the robot along the optimized motion trajectory, the position and speed information of each joint of the robot are collected by using the joint end encoder of the robot, and the acceleration information of each joint is obtained by differentiating the speed; the drive current signal of each joint of the robot is obtained through the feedback of the robot joint driver, and the estimated torque of each joint of the robot is obtained through the current-torque mapping. Specifically: τ j = k j r j I j , where: I j is the drive current of each joint of the robot, k j is the motor torque constant of each joint, r j is the transmission ratio coefficient of each joint, and τ j is the estimated torque of each joint. The values of k j and r j are provided by the robot manufacturer or identified by oneself.

[0008] The spatio-temporal graph convolutional network mentioned above includes: a graph adjacency matrix construction module and a graph network module, where: the graph adjacency matrix construction module is used to generate a robot joint graph adjacency matrix composed of a spatial connection relationship matrix and a kinematic and dynamic characteristic matrix, and the graph network module extracts spatio-temporal characteristics from the input robot joint data based on the graph adjacency matrix through spatial graph convolution and time-gated convolution.

[0009] The graph adjacency matrix construction module generates a spatial relationship matrix and a kinematic and dynamic characteristic matrix based on the spatial connection relationship and kinematic and dynamic characteristics between robot joints.

[0010] Technical effects

[0011] Based on the spatial connection relationship and kinematic and dynamic characteristics between robot joints, the present invention constructs a spatial relationship matrix and a kinematic and dynamic characteristic matrix, and further constructs a robot joint graph adjacency matrix applied to a spatio-temporal graph convolutional network, more fully extracting the influence relationship between the torque errors of each robot joint. Then, a spatio-temporal graph convolutional neural network is used to simultaneously predict and compensate the torque errors of each robot joint. Compared with the prior art, since the present invention considers the spatial relationship and physical characteristics between each robot joint, it more fully extracts the influence relationship between the torque errors of each robot joint, and has higher accuracy and more robust performance than other model-based torque compensation methods and deep learning-based torque compensation methods that only consider the time relationship; in addition, different from other methods that separately predict the torque errors of each joint, this method can simultaneously predict the torque errors of each robot joint. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a hardware architecture diagram of an embodiment;

[0013] Figure 2 It is a schematic flowchart of the present invention;

[0014] Figure 3 It is a schematic diagram of the principle of the present invention;

[0015] Figure 4 It is a diagram of the training process of an embodiment;

[0016] Figure 5 and Figure 6 It is a diagram for comparing the effects of an embodiment (Joint 1, Joint 2). DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] As Figure 1 shown, the environment setting of this embodiment includes: a system control module and a robot execution module, where: the system control module includes a host computer, an industrial PC, a communication bus, and a robot control cabinet. The host computer receives the joint data fed back by the robot as the input of the spatio-temporal graph convolutional network and outputs the predicted torque error of the robot joint. The industrial PC and the robot control cabinet communicate in real time; the robot execution module includes a servo motor and a robot body. The servo motor receives the control instruction issued by the control cabinet and executes it to make the robot body move, and performs error compensation based on the predicted torque error of the robot joint.

[0018] The robot body selected is a six-degree-of-freedom industrial robot, and its six joints are all rotary joints. Based on the real-time PLC kernel of the controller, the system can collect the state information of the servo motor at a maximum frequency of 20Khz.

[0019] As Figure 2As shown in the figure, this embodiment relates to an industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network for the above application scenarios, specifically including:

[0020] Step 1, generate the robot motion trajectory:

[0021] 1.1) Generate the excitation trajectory of the robot motion in the form of a fifth-order Fourier series, specifically: where: N = 5 is the order of the selected Fourier series, a i，k and b i，k are the coefficients of the k-th trigonometric function of the i-th joint, q i，0 is a constant term to ensure that the trajectory can meet the initial position and speed constraints, and w f is the fundamental frequency of the excitation trajectory.

[0022] The excitation trajectory in the form of a fifth-order Fourier series is used because it is differentiable multiple times, and it is easy to obtain the analytical expressions of velocity and acceleration. Moreover, the flexible effect of the robot can be avoided by designing the frequency range.

[0023] 1.2) Use the Patternsearch toolbox in MATLAB to further optimize the robot motion trajectory by the method of the number of constraint conditions to cover more motion states, and obtain an optimized motion trajectory that meets the robot joint position and speed limits through physical constraints.

[0024] The generated excitation trajectory of the robot motion can serve both the robot dynamics parameter identification task and the subsequent training task of the robot joint torque compensation network.

[0025] Step 2, collect the robot joint positions, speeds and calculate the estimated torques of each robot joint:

[0026] 2.1) When the robot moves according to the excitation trajectory obtained in Step 1, use the robot joint end encoder to collect the position and speed information of each robot joint, and use the robot joint driver feedback to obtain the drive current signal of each robot joint.

[0027] To avoid the contingency of data collection and reduce the noise in the data, perform mean filtering on the above robot joint position, speed and current data.

[0028] To avoid the huge noise in the acceleration information obtained by velocity differentiation, use a fifth-order Butterworth low-pass filter to further filter the speed information of each robot joint.

[0029] 2.2) Perform differential processing on the speed information of each robot joint obtained in Step 2.1 to obtain the acceleration information of each joint.

[0030] 2.3) Obtain the estimated torque τ of each joint of the robot through current-torque mapping j = k j r j I j , where: I j is the drive current signal of each joint of the robot obtained in step 2.1), k j is the motor torque constant of each joint, r j is the transmission ratio coefficient of each joint, and τ j is the estimated torque of each joint.

[0031] In this embodiment, the values of k j and r j are provided by the robot manufacturer or obtained by self-identification.

[0032] Step three, perform robot dynamic parameter identification, obtain a more accurate robot dynamic model, and obtain the preliminary joint prediction torque:

[0033] 3.1) Establish a robot dynamic model according to the Newton-Euler formula and linearize it, then derive the minimum inertia parameter set of the robot, and solve for the basic dynamic parameters of the robot through the identification method. Specifically: This formula can be obtained through the linearization of the robot dynamic model, where: π ∈ R m×1 is the minimum inertia parameter set of the robot to be identified, m is the number of minimum inertia parameters, τ ∈ R 6×1 is the estimated torque of the 6 joints of the robot obtained in step two, and 6 represents the number of robot joints; the observation matrix is determined by the position, velocity information and acceleration information of each joint, and a more accurate basic dynamic parameter π of the robot is obtained by solving through the weighted least squares method.

[0034] 3.2) Based on the basic dynamic parameter π of the robot, determine the corresponding observation matrix according to the position, velocity information and acceleration information of each joint of the robot at each moment. Denote the observation matrix at time s as Furthermore, obtain the preliminary prediction torque of each joint of the robot at this time from the robot inverse dynamics, that is, calculate where: τ s is the preliminary prediction torque of each joint of the robot at time s. There are still torque errors in this prediction torque caused by non-modelable non-linear and random influencing factors, and these errors are compensated by the subsequent method based on the spatio-temporal graph convolutional neural network.

[0035] Step four, construct the joint graph adjacency matrix according to the spatial relationship, kinematic and dynamic characteristics between the robot joints:

[0036] 4.1) As shown in Figure 3 the figure, 6×4 features including the position, velocity information, acceleration information, and preliminary predicted torque of each joint of the six-degree-of-freedom industrial robot in this embodiment are collected. Among them, the position, velocity information, and acceleration information are kinematic features, and the preliminary predicted torque is a dynamic feature.

[0037] 4.2) The topological graph of each joint of the robot is G=(V, E, A), where: V is the set of each joint node of the robot, E is the set of the edges where each joint node of the robot is located, and A is the adjacency matrix between each joint node of the robot, reflecting the connection relationship between each joint.

[0038] 4.3) Based on the spatial relationship between the robot joints, a preliminary joint graph adjacency matrix is constructed to reflect the physical space connection relationship between each joint. Each joint of the serial industrial robot is connected in sequence, and there is a direct spatial connection relationship between two adjacent joints, that is, the spatial relationship matrix where: 1 indicates that there is a physical connection relationship between two adjacent joints, 0 indicates that there is no physical connection relationship, and i is the above 4 different types of features. Therefore, the adjacency matrix A1 = diag[P1, P2, P3, P4] constructed according to the spatial relationship between each joint.

[0039] 4.4) Based on the physical characteristics between the robot joints, the joint graph adjacency matrix is improved to reflect the obvious physical characteristics between each joint: the kinematic information of the robot such as velocity and acceleration can be iteratively derived from the inside out, while the dynamic information of the robot such as force and torque is iteratively derived from the outside in. This shows that the front joints close to the robot base will affect the kinematic characteristics of the subsequent joints, but will not affect their dynamic characteristics. On the contrary, the subsequent joints far from the robot base will affect the dynamic characteristics of the front joints, but will not affect their kinematic characteristics. Therefore, the kinematic characteristic matrix and the dynamic characteristic matrix are set. Where: 1 indicates that there is a kinematic or dynamic influence relationship between two joints, 0 indicates that there is no kinematic or dynamic influence relationship between two joints, j is the above 3 different types of kinematic features, and k is the above 1 type of kinematic feature. Therefore, the adjacency matrix A2 = diag[Q1, Q2, Q3, Q4] constructed according to the physical characteristics between each joint.

[0040] 4.5) Considering the spatial relationship and physical characteristics between the above-mentioned robot joints comprehensively, the adjacency matrix A of each joint graph of the robot applied to the subsequent spatio-temporal graph neural network is constructed as A = αA1 + βA2, where: A1 is the adjacency matrix constructed according to the spatial relationship between each joint, A2 is the adjacency matrix constructed according to the physical characteristics between each joint, and α and β are the influence weight coefficients of the two relationships.

[0041] In this embodiment, the influence weight coefficients α = 1 and β = 0.5 are selected according to the experimental results. That is, taking the physical connection situation between each joint as the main factor and the kinematic and dynamic influence characteristics between joints as the auxiliary factor can more accurately describe the actual situation between each joint of the robot.

[0042] Step 5: Construct a spatio-temporal graph convolutional network including a graph adjacency matrix construction module and a graph network module in the PyTorch environment. Specifically: In order to take into account the temporal relationship and spatial relationship between the joint features in the robot torque tracking task, combine the spatial graph convolutional module with the temporal gated convolutional module to construct a spatio-temporal graph convolutional network. Specifically: Where: v l and v l+1 are the robot joint states at times l and l + 1 respectively, and are two temporal convolutional kernels, Θ l is the spatial convolutional kernel, * τ is the temporal convolution operation, is the spatial convolution operation, and ReLU is the network activation function.

[0043] The described graph adjacency matrix construction module reflects the process of describing the joint relationship of the robot, including a spatial connection relationship unit and a kinematic and dynamic characteristic unit, and reflects the relationship between the robot joint data input to the network.

[0044] The described graph network module includes a spatial graph convolutional unit and a temporal gated convolutional unit, which can specifically extract the spatio-temporal characteristics of the robot joint data.

[0045] Step 6: Preprocess the collected data to construct a data set: Filter the robot joint position, speed information, acceleration information, and joint estimated torque obtained in Step 2, and the preliminary joint prediction torque obtained in Step 3, and then construct a data set after normalization processing.

[0046] The described filtering process means: Adopt a fifth-order Butterworth low-pass filter to ensure removing noise while ensuring data accuracy as much as possible.

[0047] In the described data set, the joint position, joint speed, joint acceleration, and preliminary joint prediction torque of each joint of the robot at 20 moments before the current moment are used as the input data of the graph network model; the joint compensation torque at the current moment, that is, the difference between the preliminary joint prediction torque and the joint estimated torque, is used as the output data of the graph network model. For the data set collected over a period of time, it is divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%.

[0048] Step 7: Train a spatio-temporal graph convolutional network based on the dataset obtained in Step 6. First, select the RMSProp optimizer to train the spatio-temporal graph convolutional network on 70% of the training set. The loss function for network training where: N is the number of dataset samples, τ ij is the estimated torque of the i-th joint of the robot at the j-th time point, is the predicted torque after compensation of the i-th joint of the robot at the j-th time point. The loss function reflects the mean of the symmetric mean absolute percentage error (SMAPE) of the torques of each joint of the robot. The training results are as Figure 4 shown.

[0049] Step 8: Verify the effectiveness of the proposed robot joint torque compensation on the verification trajectory. In the online phase, select another trajectory different from the robot motion excitation trajectory in Step 1 as the verification trajectory, and then use the trained spatio-temporal graph convolutional network in Step 7 for real-time joint torque compensation.

[0050] The experimental results show that the torque errors of the six joints of the robot are effectively compensated, and are significantly better than the model-based method and the methods based on time series networks such as TCN.

[0051] The torque compensation result of Joint 1 is as Figure 5 shown. The RMSE of the predicted torque error based on the model is 5.69, the MAE is 4.45, and the SMAPE is 0.1110. The RMSE of the predicted torque error based on the TCN network is 4.23, the MAE is 3.30, and the SMAPE is 0.0912. The RMSE of the predicted torque error based on the STGCN network proposed in the present invention is 2.56, the MAE is 2.12, and the SMAPE is 0.0634.

[0052] The torque compensation result of Joint 2 is as Figure 6 shown. The RMSE of the predicted torque error based on the model is 18.49, the MAE is 10.18, and the SMAPE is 0.2248. The RMSE of the predicted torque error based on the TCN network is 10.55, the MAE is 7.38, and the SMAPE is 0.2160. The RMSE of the predicted torque error based on the STGCN network proposed in the present invention is 3.80, the MAE is 2.88, and the SMAPE is 0.1279. The compensation results of the remaining joints are similar and have all been significantly improved. Among them, RMSE is the Root Mean Square Error, MAE is the Mean Absolute Error, and SMAPE is the Symmetric Mean Absolute Percentage Error.

[0053] Compared with the prior art, the present method constructs a spatio-temporal graph convolutional network based on the graph adjacency matrix, which can comprehensively consider the spatial relationship and physical characteristics between the robot joints, so as to achieve higher-precision prediction and compensation of the torque errors of the six joints of the robot. Compared with the model-based prediction method and other deep learning methods that only consider the time relationship, the present invention can achieve better torque compensation effect. After compensation, the performance indexes of torque error of each joint, namely RMSE, MAE and SMAPE, have been significantly improved.

[0054] Those skilled in the art can make local adjustments to the above specific implementation in different ways without departing from the principle and purpose of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation, and all implementation solutions within its scope are subject to the present invention.

Claims

1. An industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network, characterized in that In the offline stage, an optimized motion trajectory that conforms to the robot joint position and speed limits is generated. The robot is set to move along the optimized motion trajectory, and the robot joint positions, speeds, and joint currents used to calculate the estimated torques of each robot joint are synchronously collected. The robot dynamics model is obtained through robot dynamics parameter identification, and then the preliminary joint prediction torques are obtained. According to the spatial relationship between robot joints and kinematic and dynamic characteristics, a joint graph adjacency matrix is constructed, thereby obtaining a sample set for training the spatio-temporal graph convolutional network; In the online stage, the torque compensation amounts of each joint at each moment are predicted using the trained spatio-temporal graph convolutional network, that is, the difference between the preliminary joint prediction torques and the joint estimated torques; This compensated torque is fed forward to the driving motors of each robot joint, thereby effectively reducing the torque tracking error of each robot joint; The spatio-temporal graph convolutional network described above includes: a graph adjacency matrix construction module and a graph network module, where: the graph adjacency matrix construction module is used to generate a robot joint graph adjacency matrix composed of a spatial connection relationship matrix and a kinematic and dynamic characteristic matrix, and the graph network module extracts spatio-temporal characteristics from the input robot joint data based on the graph adjacency matrix through spatial graph convolution and time-gated convolution.

2. The industrial robot joint torque compensation method based on the spatio-temporal graph convolutional network according to claim 1, characterized in that The optimized motion trajectory uses an excitation trajectory in the form of a fifth-order Fourier series as the training trajectory for robot dynamics parameter identification and torque compensation. After optimizing the robot motion trajectory through the condition number of constraints, the optimized motion trajectory that conforms to the robot joint position and speed limits is obtained through physical constraints.

3. The method for compensating the joint torque of an industrial robot based on a spatio-temporal graph convolutional network according to claim 1, wherein The acquisition mentioned above refers to: during the movement of the robot along the optimized motion trajectory, using the joint - end encoders of the robot to collect the position and velocity information of each joint of the robot, and obtaining the acceleration information of each joint through velocity differentiation; using the feedback of the robot joint drivers to obtain the drive current signals of each joint of the robot, and obtaining the estimated torque of each joint of the robot through current - torque mapping. Specifically: , where: is the drive current of each joint of the robot, is the motor torque constant of each joint, is the transmission ratio coefficient of each joint, is the estimated torque of each joint, and The values of are provided by the robot manufacturer or obtained through self - identification.

4. The industrial robot joint torque compensation method based on the spatio-temporal graph convolutional network according to any one of claims 1-3, characterized in that specifically Including: Step one, generate the robot motion trajectory: 1.1) Generate the excitation trajectory of the robot's motion in the form of a fifth-order Fourier series, specifically: , where: is the selected order of the Fourier series, and are the coefficients of the -th trigonometric function for the -th joint, is a constant term to ensure that the trajectory can satisfy the initial position and velocity constraints, is the fundamental frequency of the excitation trajectory; 1.2) Use the Patternsearch toolbox in MATLAB to further optimize the robot motion trajectory through the method of the condition number of constraints to cover more motion states, and obtain the optimized motion trajectory that conforms to the robot joint position and speed limits through physical constraints; Step two, collect the robot joint positions, speeds and calculate the estimated torques of each robot joint; Step three, perform robot dynamics parameter identification to obtain a relatively accurate robot dynamics model and obtain the preliminary joint prediction torques; Step four, construct a joint graph adjacency matrix according to the spatial relationship between robot joints and kinematic and dynamic characteristics; Step 5: Construct a spatio-temporal graph convolutional network including a graph adjacency matrix construction module and a graph network module in the PyTorch environment. Specifically: To take into account the temporal and spatial relationships between joint features in the robot torque tracking task, combine the spatial graph convolutional module with the temporal gated convolutional module to construct a spatio-temporal graph convolutional network, specifically: , where: and are respectively and the robot joint states at moments and are two temporal convolutional kernels, is the spatial convolutional kernel, is the temporal convolution operation, is the spatial convolution operation, is the network activation function; Step six, preprocess the collected data and construct a data set: filter the robot joint position, speed information, acceleration information, and joint estimated torques obtained in step two, and the preliminary joint prediction torques obtained in step three, and then construct a data set after normalization processing; Step 7: Train a spatio-temporal graph convolutional network based on the dataset obtained in Step 6. First, use the RMSProp optimizer to train the spatio-temporal graph convolutional network on 70% of the training set. The loss function for network training is , where: is the number of samples in the dataset, is the estimated torque of the -th joint of the robot at the -th time point, is the compensated predicted torque of the -th joint of the robot at the -th time point; the loss function reflects the mean of the symmetric mean absolute percentage errors of the torques of each joint of the robot. Step eight, verify the effectiveness of the proposed robot joint torque compensation on the verification trajectory: In the online stage, select another trajectory different from the robot motion excitation trajectory in step one as the verification trajectory, and then use the spatio-temporal graph convolutional network trained in step seven for real-time joint torque compensation.

5. The industrial robot joint torque compensation method based on a spatio-temporal graph convolutional network according to claim 4, characterized in that The specific content of step two includes: 2.1) When the robot moves along the excitation trajectory obtained in Step 1, the position and velocity information of each joint of the robot are collected using the joint - end encoder of the robot, and the drive current signal of each joint of the robot is obtained through the feedback of the robot joint driver; 2.2) The velocity information of each joint of the robot obtained in Step 2.1 is processed by differentiation to obtain the acceleration information of each joint; 2.3) Obtain the estimated torque of each joint of the robot through current-torque mapping , where: is the drive current signal of each joint of the robot obtained in step 2.1), is the motor torque constant of each joint, is the transmission ratio coefficient of each joint, is the estimated torque of each joint.

6. The method for compensating the joint torque of an industrial robot based on a spatio-temporal graph convolutional network according to claim 4, characterized in that Step 3 specifically includes: 3.1) Establish the robot dynamic model according to Newton-Euler formula and linearize it, then derive the minimum inertia parameter set of the robot, and solve the basic dynamic parameters of the robot through the identification method. Specifically: , where: is the minimum inertia parameter set of the robot to be identified, is the number of minimum inertia parameters, is the estimated torque of the 6 joints of the robot obtained in step 2; the observation matrix is determined by the position, velocity information and acceleration information of each joint, and the relatively accurate basic dynamic parameters of the robot are obtained by solving through the weighted least squares method ; 3.2) Based on the basic dynamic parameters of the robot , according to the position, velocity information and acceleration information of each joint of the robot at each moment, determine the corresponding observation matrix, denoted as the observation matrix at time , and then obtain the preliminary predicted torque of each joint of the robot at this time from the inverse dynamics of the robot, that is, calculate , where: is the preliminary predicted torque of each joint of the robot at time 7. The method for compensating the joint torque of an industrial robot based on a spatio-temporal graph convolutional network according to claim 4, wherein Step 4 specifically includes: 4.1) Collect 6×4 features including the position, velocity information, acceleration information, and preliminary predicted torque of each joint of the six - degree - of - freedom industrial robot. Among them, the position, velocity information, and acceleration information are kinematic features, and the preliminary predicted torque is a dynamic feature; 4.2) The topological graph of each joint of the robot is , where: is the set of each joint node of the robot, is the set of the edges where each joint node of the robot is located, is the adjacency matrix between each joint node of the robot, reflecting the connection relationship between each joint; 4.3) Construct a preliminary joint graph adjacency matrix based on the spatial relationship between the robot joints to reflect the physical spatial connection relationship between each joint; the joints of the serial industrial robot are connected in sequence, and there is a direct spatial connection relationship between two adjacent joints, that is, the spatial relationship matrix , where: means there is a physical connection relationship between two adjacent joints, means there is no physical connection relationship, are the above 4 different categories of features; therefore, the adjacency matrix constructed according to the spatial relationship between each joint ; 4.4) Improve the joint graph adjacency matrix based on the physical characteristics between robot joints to reflect the obvious physical characteristics between joints: the kinematic information of the robot, such as velocity and acceleration, can be iteratively derived from the inside out, while the dynamic information of the robot, such as force and torque, is iteratively derived from the outside in. This shows that the front joints close to the robot base will affect the kinematic characteristics of the subsequent joints, but will not affect their dynamic characteristics. On the contrary, the subsequent joints far from the robot base will affect the dynamic characteristics of the front joints, but will not affect their kinematic characteristics. Therefore, set the kinematic characteristic matrix , the dynamic characteristic matrix , where: indicates that there is a kinematic or dynamic influence relationship between two joints, indicates that there is no kinematic or dynamic influence relationship between two joints, are the above 3 different categories of kinematic characteristics, is the above 1 category of kinematic characteristics; Therefore, the adjacency matrix constructed according to the physical characteristics between joints; 4.5) Considering the above-mentioned spatial relationship and physical characteristics between the robot joints comprehensively, construct the graph adjacency matrix of each joint of the robot applied to the subsequent spatio-temporal graph neural network , where: is the adjacency matrix constructed according to the spatial relationship between the joints, is the adjacency matrix constructed according to the physical characteristics between the joints, and are the influence weight coefficients of the two relationships.

Images