Motor-driven joint characteristic modeling and compensation control method for execution error

Abstract

The application discloses a compensation control method based on motor driving joint characteristic modeling and execution error, and on the basis of joint closed-loop control, proposes a hysteresis model of the joint based on an improved iTransformer network, in modeling, taking the iTransformer network as a basic structure, improving from two aspects of structure and input end to improve the complex hysteresis model precision of the joint, then utilizing the model prediction result to feed forward compensate the joint angle set value, eliminating the influence of load change on the execution precision of the joint, and finally achieving high-precision positioning control of the industrial robot system.

Description

Technical Field

[0001] This invention relates to the field of industrial robot technology, specifically to a method for compensating for execution errors based on motor-driven joint characteristic modeling. Background Technology

[0002] In recent years, industrial robot control technology has continuously developed and progressed, finding widespread application in fields such as automotive electronics, aerospace, semiconductor chips, welding, and sorting. With the increasing demands of production and manufacturing processes, the precision requirements for industrial robots are also rising. In industrial robots, flexible joints, composed of motors and harmonic reducers, exhibit complex hysteresis characteristics, which significantly affect positioning accuracy. The hysteresis characteristics exhibited by the flexible joints of industrial robots are mainly caused by factors such as joint parameter errors, nonlinear friction, stiffness, and harmonic reducer clearance.

[0003] To improve the precision of industrial robots, high-precision joint compensation control based on hysteresis models is an important technical approach to enhance robot control accuracy. However, existing literature mainly focuses on hysteresis modeling for piezoelectric ceramics and electromagnets, which exhibit weak nonlinear hysteresis characteristics, rather than on robot joints exhibiting complex hysteresis characteristics such as abrupt changes, non-smoothness, and strong nonlinearity in both forward and reverse strokes. Therefore, directly applying existing hysteresis modeling methods to robot joint compensation control inevitably leads to errors. Furthermore, most existing literature treats joint characteristics as general nonlinear objects and directly controls them, resulting in persistent issues with low precision in joint execution angles. Summary of the Invention

[0004] The present invention aims to address the problem that the joint angle execution accuracy of industrial robots varies with load, which limits the overall control accuracy of industrial robots. It provides a compensation control method based on motor-driven joint characteristic modeling and execution error compensation.

[0005] To solve the above problems, the present invention is achieved through the following technical solution:

[0006] The method for compensating for execution errors based on motor-driven joint characteristic modeling includes the following steps:

[0007] Step 1: Construct an improved iTransformer network. This improved iTransformer network consists of an autocorrelation-type hysteresis module, an embedding layer, a TrmBlock (Transformer Block), a fully connected layer, and an LSTM (Long Short-Term Memory) network. The input of the autocorrelation-type hysteresis module forms the input of the improved iTransformer network. The output of the autocorrelation-type hysteresis module is connected to the input of the embedding layer and the LSTM network. The output of the embedding layer is connected to the input of the fully connected layer via the TrmBlock. The output of the LSTM network is directly connected to the input of the fully connected layer. The output of the fully connected layer forms the output of the improved iTransformer network.

[0008] Step 2: Collect the torque and torsion angle of the robot joints at historical moments, and then... Torque at any moment , Torque at any moment and Torque at any moment ,as well as Angle of twist at any moment , Angle of twist at any moment and Angle of twist at any moment They are fed into the improved iTransformer network, and the improved iTransformer network outputs... Predicted value of torsion angle error at time ;

[0009] Step 3: Utilize the output of the improved iTransformer network Predicted value of torsion angle error at time For the setting Setting the torsion angle at the moment Compensation is performed to obtain the compensated torsion angle. ,in ;

[0010] Step 4: Adjust the compensated torsion angle The data is sent to the control terminal of the robot joint to achieve compensatory control of the robot joint.

[0011] In step 1 above, the new input vector output by the autocorrelation-type hysteresis module is:

[0012] ;

[0013] in, Represents a class of hysteresis operators. ; Represents the autocorrelation vector. ; , , They represent time, time, Torque at any given moment; , , They represent time, time, The angle of twist at any moment.

[0014] Compared with existing technologies, this invention proposes a joint hysteresis model based on an improved iTransformer network, building upon joint closed-loop control. In the modeling process, the iTransformer network is used as the basic structure, and improvements are made to both the structure and the input to enhance the accuracy of the complex hysteresis model of the joint. Furthermore, the model prediction results are used to perform feedforward compensation on the joint angle setpoint, eliminating the impact of load variations on joint execution accuracy, ultimately achieving high-precision positioning control for the industrial robot system. Attached Figure Description

[0015] Figure 1 Here is a diagram of the iTransformer network structure;

[0016] Figure 2 This is a schematic diagram of the hysteresis model of a joint based on the iTransformer network.

[0017] Figure 3 The diagram shows the validation results of the hysteresis model of the joint based on the iTransformer network. (a) Validation results, (b) Validation error.

[0018] Figure 4 This is a basic structure diagram of an LSTM unit;

[0019] Figure 5 A schematic diagram of a hysteresis loop;

[0020] Figure 6 This is a schematic diagram of the hysteresis model of a joint based on the improved iTransformer network. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0022] (1) Hysteresis model of joints based on improved iTransformer network:

[0023] This invention proposes to use the iTransformer network as the basic structure to design a complex hysteresis model for a robot's flexible joint. The iTransformer network is used to model the hysteresis characteristics because its self-attention mechanism can focus on the most prominent information in the hysteresis characteristics, ignoring invalid information. Higher weights are used to focus on important information, while lower weights ignore invalid or interfering information, similar to the hysteresis characteristics of remembering and forgetting input information.

[0024] The iTransformer network is an improvement on the Transformer model, with a simpler architecture that uses only the encoding structure of the Transformer model. The iTransformer network includes embedding layers, TrmBlock, and fully connected layers with mapping capabilities, such as... Figure 1 As shown, the input of the embedding layer is the input of the iTransformer network, the output of the embedding layer is connected to TrmBlock, the output of TrmBlock is connected to the input of the fully connected layer, and the output of the fully connected layer forms the output of the iTransformer network.

[0025] In TrmBlock, the mathematical expression for the self-attention mechanism is:

[0026] ;

[0027] in, For the input vector, , , For bias vectors, For query vector, For key vectors, For value vectors, For the dimension of the key, For activation function, This is the output of the self-attention mechanism.

[0028] The overall mathematical expression for the iTransformer hysteresis model is:

[0029] ;

[0030] in, The first element in the model's input vector Column elements, For the input vector of the th The result obtained by embedding column elements. For the process The result obtained from layer TrmBlock , For the process The result obtained from layer TrmBlock , For the process The result obtained from layer TrmBlock The Column elements, The result obtained from the fully connected layer is the output vector of the model. In this embodiment, the number of columns of the input model vector is specified. . In this embodiment, the number of layers of TrmBlock is... .

[0031] Hysteresis models of joints based on iTransformer networks, such as those constructed using iTransformer networks. Figure 2 As shown, the joint time, time, Torque at any moment , , and time, time, Angle of twist at any moment , , They are all fed into the iTransformer network and used to make predictions. Predicted value of torsion angle error at time .

[0032] Figure 3 The figures show the validation results of the hysteresis model of a joint based on the iTransformer network: (a) validation results, (b) validation error. Experiments show that directly using the iTransformer network to model the hysteresis characteristics of joints is not effective, resulting in significant model errors. Therefore, this invention improves the iTransformer network to enhance the accuracy of the hysteresis model.

[0033] To further improve the accuracy of hysteresis models, this invention proposes an improved iTransformer network, which mainly consists of an autocorrelation-based hysteresis module, an embedding layer, a TrmBlock, a fully connected layer, and an LSTM network. The input of the autocorrelation-based hysteresis module forms the input of the improved iTransformer network. The output of the autocorrelation-based hysteresis module is connected to the inputs of the embedding layer and the LSTM network. The output of the embedding layer is connected to the input of the fully connected layer via the TrmBlock. The output of the LSTM network is directly connected to the input of the fully connected layer. The output of the fully connected layer forms the output of the improved iTransformer network.

[0034] 1) Autocorrelation-based hysteresis modules:

[0035] By designing an autocorrelation-type hysteresis module, the correlation of hysteresis characteristics between input vectors is strengthened. The input data is processed and an autocorrelation vector is constructed. The autocorrelation vector is then combined with the hysteresis-type operator to obtain a new input vector.

[0036] Since torsion angles often exhibit continuity and correlation over time, the product of the torsion angle at a previous moment and the torsion angles at other moments can reflect the degree of correlation between them. To enhance the iTransformer network's ability to extract and represent features from the torsion angle sequence in the input data, this invention utilizes the Hadamard Product to enhance the autocorrelation and feature strengthening of the torsion angle sequence. A new torsion angle autocorrelation vector is generated by multiplying the input torsion angle at the previous moment vectorwise with the torsion angles at other moments. Its mathematical expression is:

[0037] ;

[0038] in, It is Hadamaji.

[0039] The new vector generated by the Hadamard product operation not only retains the information of the original twist angle but also incorporates the interactive features between twist angles at different times.

[0040] Design a simple hysteresis-like operator. Its mathematical expression is:

[0041] ;

[0042] Hysteresis operators and Angle of twist at any moment Constructing a hysteresis loop, such as Figure 4 As shown.

[0043] Hysteresis-like feature information is obtained from the input signal and used as the input front-end information of the iTransformer network, which enhances its ability to describe hysteresis characteristics.

[0044] Combining the autocorrelation vector with a hysteresis-like operator yields a new input vector, the mathematical expression of which is:

[0045] ;

[0046] 2) LSTM network:

[0047] The basic structure of an LSTM unit is as follows: Figure 5 As shown, its mathematical model is as follows:

[0048] ;

[0049] in, This represents the input at time t. and These represent the outputs of the hidden layer LSTM cells at time t-1 and time t, respectively. express Activation function; express Activation function; , , , , , , , and , , , These represent the input gate, forget gate, output gate, and current input cell state, respectively. The weight and bias matrix; Represents the pointwise multiplication of the Hadamard product; Input gate Used to control the proportion of input stored in memory at the current moment; forget gate The output gate determines the proportion of the value of the memory cell from the previous time step that is output to the current time step. The proportion of values ​​stored in a control memory unit that can be output; memory unit It is a weighted sum of the previous time value and the current input cell state.

[0050] The hysteresis model of the joints based on the improved iTransformer network constructed using the above-mentioned improved iTransformer network is as follows: Figure 6 As shown, the torque of the joint is... time, time, Torque at any moment , , and time, time, Angle of twist at any moment , , They are fed into the improved iTransformer network and used to make predictions. Predicted value of torsion angle error at time The joint output was obtained based on measurements. and difference The parameters in the hysteresis model of the joints of the improved iTransformer network are updated using the steepest descent method.

[0051] The characteristics of the hysteresis model for joints based on the improved iTransformer network are as follows: After the vector is mapped to a new feature space through an embedding layer with an inversion effect, the correlation between different embedding vectors and the hysteretic static nonlinear features are captured by the self-attention mechanism. The self-attention mechanism can adaptively focus on the input information at different times and extract key features. The constructed autocorrelation vector is processed by an LSTM network. The autocorrelation vector enhances the LSTM's ability to capture long-term dependencies in hysteresis. Through a gating mechanism, information is selectively remembered and forgotten, thereby modeling the dynamic behavior of hysteresis systems. The structural improvements and the introduction of autocorrelation-type hysteresis modules can better capture the static and dynamic nonlinear characteristics of hysteresis. The improved model mainly constructs an autocorrelation vector that reflects the hysteresis characteristics through autocorrelation-type hysteresis modules.

[0052] (2) Error compensation control of joints based on feedforward control:

[0053] The torque value, which varies with the load, is obtained by a torque sensor installed inside the joint. At time t, the torque reflects the load magnitude. As the input signal, the predicted value of the joint's execution angle error is obtained through the output of the invented hysteresis model. The technical term is also torsion angle. Feedforward compensation is performed before the joint executes the angle command; the current joint angle command is... The command to execute the compensated joint angle is as follows: The compensation formula is:

[0054]

[0055] After compensation, the set angle value can be modified online from the joint input command terminal, indirectly reducing the impact of load changes on the joint angle execution accuracy. By compensating and controlling each joint of the robot, the impact of corresponding load changes on control accuracy can be effectively suppressed or eliminated under different postures and loads, achieving the goal of high-precision robot control.

[0056] Based on the above analysis, the present invention proposes a method for compensating for execution errors based on motor-driven joint characteristic modeling, which includes the following steps:

[0057] Step 1: Construct an improved iTransformer network. This improved iTransformer network consists of an autocorrelation-type hysteresis module, an embedding layer, a TrmBlock, a fully connected layer, and an LSTM network. The input of the autocorrelation-type hysteresis module forms the input of the improved iTransformer network. The output of the autocorrelation-type hysteresis module is connected to the input of the embedding layer and the LSTM network. The output of the embedding layer is connected to the input of the fully connected layer via the TrmBlock. The output of the LSTM network is directly connected to the input of the fully connected layer. The output of the fully connected layer forms the output of the improved iTransformer network.

[0058] The new input vector output by the autocorrelation-type hysteresis module is:

[0059] ;

[0060] in, Represents a class of hysteresis operators. ; Represents the autocorrelation vector. ; , , They represent time, time, Torque at any given moment; , , They represent time, time, The angle of twist at any moment.

[0061] Step 2: Collect the torque and torsion angle of the robot joints at historical moments, and then... Torque at any moment , Torque at any moment and Torque at any moment ,as well as Angle of twist at any moment , Angle of twist at any moment and Angle of twist at any moment They are fed into the improved iTransformer network, and the improved iTransformer network outputs... Predicted value of torsion angle error at time .

[0062] Step 3: Utilize the output of the improved iTransformer network Predicted value of torsion angle error at time For the setting Setting the torsion angle at the moment Compensation is performed to obtain the compensated torsion angle. ,in .

[0063] Step 4: Adjust the compensated torsion angle The data is sent to the control terminal of the robot joint to achieve compensatory control of the robot joint.

[0064] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A method for compensating for execution errors based on motor-driven joint characteristic modeling, characterized by: The steps include the following: Step 1: Construct an improved iTransformer network. This improved iTransformer network consists of an autocorrelation-type hysteresis module, an embedding layer, a transformation module, a fully connected layer, and an LSTM network. The input of the autocorrelation-type hysteresis module forms the input of the improved iTransformer network. The output of the autocorrelation-type hysteresis module is connected to the input of the embedding layer and the LSTM network. The output of the embedding layer is connected to the input of the fully connected layer via the transformation module. The output of the LSTM network is directly connected to the input of the fully connected layer. The output of the fully connected layer forms the output of the improved iTransformer network. The new input vector output by the autocorrelation-type hysteresis module is: ； in, Represents a class of hysteresis operators. ; Represents the autocorrelation vector. ; , , They represent time, time, Torque at any given moment; , , They represent time, time, The angle of twist at any moment; Step 2: Collect the torque and torsion angle of the robot joints at historical moments, and then... Torque at any moment , Torque at any moment and Torque at any moment ,as well as Angle of twist at any moment , Angle of twist at any moment and Angle of twist at any moment They are fed into the improved iTransformer network, and the improved iTransformer network outputs... Predicted value of torsion angle error at time ; Step 3: Utilize the output of the improved iTransformer network Predicted value of torsion angle error at time For the setting Setting the torsion angle at the moment Compensation is performed to obtain the compensated torsion angle. ,in ; Step 4: Adjust the compensated torsion angle The data is sent to the control terminal of the robot joint to achieve compensatory control of the robot joint.

Images