Joint module hysteresis modeling and compensation method of driving motor and speed reducer

Abstract

The application discloses a hysteresis modeling and compensation method for a joint module composed of a driving motor and a speed reducer, and comprises the following steps: firstly, a hysteresis model based on a dynamic hyperbolic tangent function and a hybrid expert neural network is constructed; secondly, joint load torque and an actual joint torsion angle of a collaborative robot joint module composed of a driving motor and a speed reducer are sent into the hysteresis model to obtain a joint predicted torsion angle at the current time; and finally, the set rotation angle of the driving motor of the collaborative robot joint module is compensated by using the joint predicted torsion angle at the current time. The hysteresis model introduces a DY-T activation function, which can effectively enhance the fitting ability of the hysteresis model to the strong nonlinear characteristics of the joint. The set rotation angle of the driving motor at the input end of the joint is compensated and corrected based on the joint execution error predicted by the hysteresis model, so that the influence of load change on the execution accuracy of the joint output end is indirectly eliminated, and finally the high-precision positioning control of the collaborative robot system is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial robot technology, specifically to a method for hysteresis modeling and compensation of joint modules composed of a drive motor and a reducer. Background Technology

[0002] In recent years, collaborative robots have become one of the important pieces of equipment in intelligent manufacturing. In the joint modules of collaborative robots, which consist of drive motors and reducers, the harmonic reducer transmission, due to factors such as the periodic elasto-plastic deformation of the flexible wheel and nonlinear friction, causes the load torque and angular execution error (torsion angle) of the joint module to exhibit strong nonlinear and non-smooth hysteresis characteristics, rather than the ideal zero-error execution characteristic. This non-zero hysteresis error seriously affects the execution accuracy of the joint. Therefore, high-precision modeling of the complex hysteresis characteristics of the joint module and model-based compensation control of the joint module's execution error are effective technical approaches to improve the execution accuracy of the joint module.

[0003] In the field of hysteresis characteristic modeling, in addition to traditional methods (Preisach model, Prandtl-Ishlinskii model, Bouc-Wen model, and Maxwell model, etc.), with the development of AI technology, various hysteresis characteristic modeling methods based on AI technology have emerged. Some literature uses GRU (Gated Recurrent Unit) neural networks to model the hysteresis of piezoelectric actuator displacement; others use hysteresis models built on LSTM (Long Short-Term Memory) networks to study the hysteresis modeling of shape memory alloys; and still others use Preisach-RNN hybrid models to model tendon sheath actuators (TMS) and describe the nonlinear hysteresis characteristics of TMS. However, these models are limited to modeling the hysteresis characteristics of objects with weak nonlinear characteristics in both forward and reverse strokes, such as piezoelectric ceramics and magnetic cores, and are not suitable for modeling the hysteresis characteristics of collaborative robot joint modules. Summary of the Invention

[0004] The present invention aims to solve the problem of complex hysteresis characteristics caused by the inherent mechanical structure of the joint module of a collaborative robot composed of a drive motor and a reducer, and provides a method for hysteresis modeling and compensation of the joint module composed of a drive motor and a reducer.

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

[0006] The method for hysteresis modeling and compensation of joint modules composed of drive motors and reducers includes the following steps:

[0007] Step 1: Construct a hysteresis model based on the dynamic hyperbolic tangent function and a hybrid expert neural network. This hysteresis model includes a preprocessing module, an improved hybrid expert neural network, and a dynamic hyperbolic tangent function feedforward neural network. The improved hybrid expert neural network consists of a gating network and a sub-expert neural network, where the sub-expert neural network is a dynamic hyperbolic tangent function feedforward neural network. The input of the preprocessing module forms the input of the hysteresis model, the output of the preprocessing module is connected to the input of the improved hybrid expert neural network, the output of the improved hybrid expert neural network is connected to the input of the dynamic hyperbolic tangent function feedforward neural network, and the output of the dynamic hyperbolic tangent function feedforward neural network forms the output of the hysteresis model.

[0008] Step 2: Calculate the joint load torque of the collaborative robot's joint module at the previous moment. Joint load torque at the first two moments The actual torsion angle of the joint at the previous moment The actual joint torsion angle at the first two moments The values ​​are fed into the hysteresis model established in step 1 to obtain the predicted joint torsion angle at the current moment. ;

[0009] Step 3: Predict the torsion angle using the joint angle obtained in Step 2 at the current moment. Setting the rotation angle of the drive motors for the joint modules of the collaborative robot Compensation is performed, and the rotation angle is set after compensation. ,in The reduction ratio of the reducer in the joint module of a collaborative robot. This represents the desired output angle of the joints in the collaborative robot's joint module at the current moment. This provides the predicted joint torsion angle for the collaborative robot's joint module at the current moment.

[0010] In step 1 above, the dynamic hyperbolic tangent function feedforward neural network is a three-layer feedforward neural network, whose activation function is a dynamic hyperbolic tangent function with three adjustable parameters.

[0011] In step 1 above, the normalization formula used by the preprocessing module is:

[0012]

[0013] In the formula, The input vector is the normalized form. For the input vector, This is the scaling factor. The mean of the input vector. The variance of the input vector. It is a constant. This is a bias term.

[0014] The actual torsional angle of a joint is equal to the expected output angle of the joint minus the actual output angle of the joint.

[0015] Compared with the prior art, the present invention has the following characteristics:

[0016] 1. In the hysteresis model, the multiple sub-expert networks of the MOE neural network correspond to different slope characteristics of the forward and reverse strokes of the non-smooth hysteresis characteristics. The gate network of the MOE neural network dynamically activates sub-expert networks with different slopes based on the input vector information, thereby realizing the direct description of the non-smooth characteristics of the joint module at the extreme point.

[0017] 2. In the hysteresis model, the MOE neural network is improved by introducing the DY-T activation function into the sub-expert neural network of the MOE neural network to realize flexible translation of the curve to adapt to the translational changes of the forward and reverse strokes of the hysteresis characteristics. At the same time, the feedforward neural network also introduces the DY-T activation function to enhance the fitting ability of the hysteresis model to the strong nonlinear characteristics of the joint.

[0018] 3. Using the designed high-precision hysteresis model, the rotation angle compensation of the joint drive motor is set at the input end of the joint module drive motor, which indirectly eliminates the execution error of the joint at the output end, and ultimately improves the execution accuracy of the joint module. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the MOE neural network.

[0020] Figure 2 This is a schematic diagram of the hysteresis model based on the MOE neural network.

[0021] Figure 3 To improve the schematic diagram of the MOE neural network.

[0022] Figure 4 This is a schematic diagram of the DY-T feedforward neural network.

[0023] Figure 5 This is a schematic diagram of the DYT-MOE hysteresis model. Detailed Implementation

[0024] 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.

[0025] This invention addresses the complex hysteresis characteristics exhibited by the joint module composed of a drive motor and a reducer, which involves strong nonlinear forward and reverse strokes due to load magnitude and execution error (torsion angle). The hysteresis is equal to the desired output angle of the joint minus the actual output angle. This invention proposes a hysteresis model based on a Dynamic Hyperbolic Function (DY-T) and a Mixture Of Experts (MOE) neural network, and uses this hysteresis model to achieve compensatory control of the joint execution error.

[0026] See Figure 1 The MOE neural network consists of a gated network (Router) and sub-expert neural networks based on FNN (Feedforward Neural Network). The gated network maps the input feature vector to a feature space of the same dimension as the number of sub-expert neural networks, and then selects the highest-scoring sub-expert neural networks. Each sub-expert neural network performs feature extraction while discarding other unselected sub-expert networks. This is a set value. Analysis revealed that the MOE neural network can instantly switch between sub-expert neural network combinations for tracking and matching when the input feature vector pattern changes, demonstrating the MOE neural network's ability to quickly capture complex feature patterns.

[0027] The mathematical description of the MOE neural network is as follows:

[0028]

[0029] In the formula, For the input feature vector, , For weighted parameters, , The dimension of the feature vector. Indicates the number of sub-expert neural networks; This means mapping the input vector to the sub-expert neural network space. This indicates taking the corresponding first [expert] from all sub-expert neural networks. The largest , This represents the normalized exponential function; Representing the The weights calculated by a gating network from a sub-expert neural network. Representing the The output of a sub-expert neural network This is the output of the MOE neural network.

[0030] The load torque and execution error torsional angle of the robot joint module exhibit hysteresis characteristics where the forward and reverse strokes do not coincide. The slopes of the forward and reverse strokes differ, displaying non-smooth characteristics at extreme points. Analysis of the MOE neural network reveals that its dynamic switching capability precisely corresponds to the non-smooth characteristics exhibited by the robot joint at these extreme points. Through a gating network mechanism that switches sub-expert combinations in real time, the MOE neural network can dynamically adapt to the discontinuous changes in the derivative of the hysteresis characteristic curve. This dynamic matching capability accurately captures the discontinuous jumps in the slope of the hysteresis curve. Therefore, the hysteresis model based on the MOE neural network differs from hysteresis models that approximate non-smooth characteristics; instead, it directly addresses the problem of abrupt slope changes in hysteresis non-smoothness.

[0031] Hysteresis models based on MOE neural networks, such as Figure 2 As shown in the figure Representing the current moment The moment before The first two moments Joint load torque; Representing the current moment The moment before The first two moments The actual torsional angle of the joint. (By...) The input vector is processed by the MOE neural network hysteresis model to obtain the joint prediction torsion angle. Predicting torsion angle based on joints With respect to the actual torsion angle of the joint The difference can be used to learn the parameters of the MOE neural network hysteresis model using the gradient method. In the MOE neural network-based hysteresis model, multiple sub-expert networks correspond to different slope characteristics of the forward and reverse strokes of the non-smooth hysteresis feature. The gating network dynamically activates sub-expert networks with different slopes based on the input vector information.

[0032] Although the hysteresis model based on the MOE neural network can directly describe the non-smooth characteristics of the joint module at extreme points, experimental data shows that while the MOE model has high fitting accuracy at the extreme points of the hysteresis loop and can describe the hysteresis characteristics, it exhibits certain translational biases in the strongly nonlinear forward and reverse stroke regions, requiring further improvement. Therefore, the hysteresis model based on the MOE neural network is improved as follows:

[0033] Improved MOE Neural Network

[0034] See Figure 3 The improved MOE neural network changes the activation function of each sub-expert neural network from the linear activation function (ReLU) to the dynamic hyperbolic tangent function (DY-T).

[0035] The mathematical description of the dynamic hyperbolic tangent function is as follows:

[0036]

[0037] Due to the parameters of the dynamic hyperbolic tangent function (DY-T) , and This allows the hyperbolic tangent function to be arbitrarily scaled, flipped, and translated, thus providing a stronger nonlinear descriptive capability than the hyperbolic tangent function. Therefore, this invention replaces the ReLU activation function of each sub-expert neural network with the DY-T activation function, enabling it to accurately fit non-smooth, nonlinear hysteresis characteristics.

[0038] The improved MOE neural network consists of a gating network and sub-expert neural networks. The gating network maps the input feature vector to a feature space of the same dimension as the number of sub-expert neural networks, and then selects the top-scoring sub-experts. Each sub-expert neural network performs feature extraction while discarding other unselected sub-expert networks. The sub-expert neural network is a Dynamic Hyperbolic Tangent Function Feedforward Neural Network (DY-T Feedforward Neural Network), a three-layer feedforward neural network structure consisting of an input layer, hidden layers, and an output layer. Its activation function is a dynamically hyperbolic tangent function with three adjustable parameters. Introducing DY-T into the MOE neural network allows for flexible curve translation to adapt to the translational changes in both forward and reverse directions of the hysteresis characteristic.

[0039] Add DY-T feedforward neural network

[0040] To further enhance the fitting ability of the hysteresis model to the strongly nonlinear characteristics of joints, this invention adds a DY-T feedforward neural network to the back end of the improved MOE neural network. See also Figure 4 The DY-T feedforward neural network is a three-layer feedforward neural network structure consisting of an input layer, a hidden layer, and an output layer. Its activation function is a dynamic hyperbolic tangent function with three adjustable parameters.

[0041] Add a preprocessing module

[0042] Because the DY-T activation function differs from the ReLU activation function, the maximum output value of the DY-T activation function is... 1. Considering that joint outputs only have positive values, and the positive part of the ReLU activation function has no amplitude limiting, the input information needs to be normalized. Therefore, this invention adds a preprocessing module to the front end of the improved MOE neural network to construct and normalize the input vector.

[0043] The input vector is from the previous time step The first two moments The joint load torque and the actual torsional angle of the joint constitute the total torque. These historical data can describe the dynamic changes in hysteresis characteristics and improve model accuracy.

[0044] The normalization formula for the input vector is as follows:

[0045]

[0046] In the formula, The input vector is the normalized form. For the input vector, This is the scaling factor. The mean of the input vector. The variance of the input vector. It is a very small constant to avoid the denominator being zero. This is a bias term.

[0047] The DYT-MOE hysteresis model constructed based on the above improvements is as follows: Figure 5 As shown. The improved MOE neural network serves as the core of the entire hysteresis model. Its gating network dynamically switches between sub-expert networks with different slopes based on abrupt changes in the input vector pattern, enabling feature extraction of the strongly non-smooth characteristics of the hysteresis curve. DY-T, as the activation function of the sub-expert neural network, enhances the ability to describe strongly nonlinear hysteresis characteristics. The combined effect of the gating network and the DY-T activation function gives the improved MOE neural network the ability to accurately fit strongly non-smooth nonlinear hysteresis characteristics. The DY-T feedforward neural network replaces the ReLU activation function with DY-T. DY-T, with its flexible translation, adapts to the translational changes in the forward and reverse strokes of the hysteresis characteristics, thereby enhancing the fitting ability of the hysteresis model to the strong nonlinear characteristics of the joints. The preprocessing module, combined with improvements to the activation function, effectively increases the learning speed of model parameters while simultaneously improving model accuracy. The learning of the DYT-MOE hysteresis model above is based on the joint prediction of torsional angles. With respect to the actual torsion angle of the joint The difference is used to obtain the hysteresis model parameters of the joint module through the gradient algorithm, thus completing the establishment of the hysteresis model.

[0048] Based on the established DYT-MOE hysteresis model, this invention compensates for the set rotation angle of the joint drive motor at the input end of the joint module composed of a drive motor and a reducer, indirectly eliminating execution errors at the joint output end and improving the execution accuracy of the joint module. In the collaborative robot joint module composed of a drive motor, reducer, etc., the reduction ratio of the reducer is set to... Ideally, the output of the joint module is , The drive motor is set to rotate at a certain angle. This represents the ideal output of the joint. However, in reality, collaborative robot joint modules will have execution errors. These execution errors can be predicted one step ahead using the established hysteresis model. In the execution error compensation, based on the model's prediction of the joint execution error, the rotation angle of the drive motor at the joint input end is set to compensate and correct, indirectly eliminating the impact of load changes on the execution accuracy of the joint output end. Through the compensation of execution errors of each joint of the robot, high-precision positioning control of the collaborative robot system is ultimately achieved.

[0049] Accordingly, this invention proposes a method for hysteresis modeling and compensation of a joint module composed of a drive motor and a reducer, comprising the following steps:

[0050] Step 1: Construct a hysteresis model based on the dynamic hyperbolic tangent function and a hybrid expert neural network. This hysteresis model includes a preprocessing module, an improved hybrid expert neural network, and a dynamic hyperbolic tangent function feedforward neural network. The improved hybrid expert neural network consists of a gating network and a sub-expert neural network, where the sub-expert neural network is a dynamic hyperbolic tangent function feedforward neural network. The input of the preprocessing module forms the input of the hysteresis model, and the output of the preprocessing module is connected to the input of the improved hybrid expert neural network. The output of the improved hybrid expert neural network is connected to the input of the dynamic hyperbolic tangent function feedforward neural network, and the output of the dynamic hyperbolic tangent function feedforward neural network forms the output of the hysteresis model.

[0051] Step 2: Calculate the joint load torque of the collaborative robot's joint module at the previous moment. Joint load torque at the first two moments The actual torsion angle of the joint at the previous moment The actual joint torsion angle at the first two moments The values ​​are fed into the hysteresis model established in step 1 to obtain the predicted joint torsion angle at the current moment. .

[0052] Step 3: Predict the torsion angle using the joint angle obtained in Step 2 at the current moment. Setting the rotation angle of the drive motors for the joint modules of the collaborative robot Compensation is performed, and the rotation angle is set after compensation. ,in The reduction ratio of the reducer in the joint module of a collaborative robot. This represents the desired output angle of the joints in the collaborative robot's joint module at the current moment. This provides the predicted joint torsion angle for the collaborative robot's joint module at the current moment.

[0053] 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 hysteresis modeling and compensation of joint modules composed of drive motors and reducers, characterized in that, The steps include the following: Step 1: Construct a hysteresis model based on the dynamic hyperbolic tangent function and a hybrid expert neural network. This hysteresis model includes a preprocessing module, an improved hybrid expert neural network, and a dynamic hyperbolic tangent function feedforward neural network. The improved hybrid expert neural network consists of a gating network and a sub-expert neural network, where the sub-expert neural network is a dynamic hyperbolic tangent function feedforward neural network. The input of the preprocessing module forms the input of the hysteresis model, the output of the preprocessing module is connected to the input of the improved hybrid expert neural network, the output of the improved hybrid expert neural network is connected to the input of the dynamic hyperbolic tangent function feedforward neural network, and the output of the dynamic hyperbolic tangent function feedforward neural network forms the output of the hysteresis model. Step 2: Calculate the joint load torque of the collaborative robot's joint module at the previous moment. Joint load torque at the first two moments The actual torsion angle of the joint at the previous moment The actual joint torsion angle at the first two moments The values ​​are fed into the hysteresis model established in step 1 to obtain the predicted joint torsion angle at the current moment. ; Step 3: Predict the torsion angle using the joint angle obtained in Step 2 at the current moment. Setting the rotation angle of the drive motors for the joint modules of the collaborative robot Compensation is performed, and the rotation angle is set after compensation. ,in The reduction ratio of the reducer in the joint module of a collaborative robot. This represents the desired output angle of the joints in the collaborative robot's joint module at the current moment. This provides the predicted joint torsion angle for the collaborative robot's joint module at the current moment.

2. The method for hysteresis modeling and compensation of a joint module composed of a drive motor and a reducer according to claim 1, characterized in that, In step 1, the dynamic hyperbolic tangent function feedforward neural network is a three-layer feedforward neural network, whose activation function is a dynamic hyperbolic tangent function with three adjustable parameters.

3. The method for hysteresis modeling and compensation of a joint module composed of a drive motor and a reducer according to claim 1, characterized in that, In step 1, the normalization formula used by the preprocessing module is: In the formula, The input vector is the normalized form. For the input vector, This is the scaling factor. The mean of the input vector. The variance of the input vector. It is a constant. This is a bias term.

4. The method for hysteresis modeling and compensation of a joint module composed of a drive motor and a reducer according to claim 1, characterized in that, The actual torsional angle of a joint is equal to the expected output angle of the joint minus the actual output angle of the joint.

Images