Motor and reducer module dynamic convolutional neural network hysteresis modeling and compensation method
By using a dynamic convolutional neural network hysteresis model and an open-loop feedforward compensation method, the problems of describing and compensating for the non-smooth hysteresis characteristics of the motor and reducer modules are solved, thereby improving the stability and accuracy of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2026-02-24
- Publication Date
- 2026-06-09
Smart Images
Figure CN122174889A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial robot technology, specifically to a method for dynamic convolutional neural network hysteresis modeling and compensation of motor and reducer modules. Background Technology
[0002] Unlike general industrial robots, collaborative robots, with their unique characteristics of high safety and good compliance, have been widely used in industries such as intelligent manufacturing. In collaborative robots constructed from motor and reducer modules, the multi-tooth coupling of the reducer within the module, the movement of the non-circular flexible wheel, and installation errors in the complex structure cause the reducer module to exhibit a strong nonlinear, non-smooth, and complex hysteresis characteristic. This hysteresis results in a torsion angle, which is the difference between the module's expected output angle and its actual output angle, and is therefore also called hysteresis error.
[0003] Traditional hysteresis models (such as the Preisach hysteresis model, Prandtl-Ishlinskii (PI) hysteresis model, Maxwell hysteresis model, and Bouc-Wen hysteresis model) are mainly used to describe the hysteresis characteristics of piezoelectric ceramics, voice coil motors, switched reluctance SRM motors with nanoscale displacement, as well as the symmetric and piecewise weakly nonlinear hysteresis characteristics exhibited by electroactive polymeric shape memory materials (EAP). Hysteresis models based on AI technology (such as the Long Short-Term Memory (LSTM) hysteresis model, Gated Recurrent Unit (GRU) hysteresis model, and Transformer hysteresis model) do not directly consider the non-smooth characteristics of hysteresis in their modeling; instead, they approximate the non-smooth hysteresis characteristics through a large number of smoothing functions and superposition approximation. Therefore, none of the above hysteresis models are suitable for describing the non-smooth, strongly nonlinear hysteresis characteristics exhibited by motor and reducer modules. Summary of the Invention
[0004] The present invention addresses the problem that existing hysteresis models are not suitable for describing the non-smooth, strongly nonlinear hysteresis characteristics exhibited by motor and reducer modules, and provides a dynamic convolutional neural network hysteresis modeling and compensation method for motor and reducer modules.
[0005] To solve the above problems, the present invention is achieved through the following technical solution:
[0006] A method for dynamic convolutional neural network hysteresis modeling and compensation in motor and reducer modules includes the following steps:
[0007] Step 1: Construct a dynamic convolutional neural network hysteresis model. This dynamic convolutional neural network hysteresis model consists of a convolutional hysteresis operator and a dynamic RBF neural network. The input of the convolutional hysteresis operator forms the input of the dynamic convolutional neural network hysteresis model, the output of the convolutional hysteresis operator is connected to the input of the dynamic RBF neural network, and the output of the dynamic RBF neural network forms the output of the dynamic convolutional neural network hysteresis model.
[0008] Step 2: Connect the motor and the reducer module. Constant load torque The result is obtained by feeding the data into a dynamic convolutional neural network hysteresis model. Predict the torsion angle at all times ;
[0009] Step 3, utilize Predict the torsion angle at all times For motor and reducer modules The motor rotation angle is set at all times. To receive compensation Motor rotation angle after time compensation ,in:
[0010]
[0011] In the formula, This refers to the reduction ratio between the motor and the reducer module.
[0012] In step 1 above, the mathematical description of the convolution hysteresis operator is as follows:
[0013]
[0014] In the formula, The output of the convolution hysteresis operator Time-based dynamic convolution, Input to the convolution hysteresis operator Constant load torque, Input to the convolution hysteresis operator Constant load torque.
[0015] Compared with the prior art, the present invention has the following characteristics:
[0016] 1. Design a dynamic hysteresis operator with the input signal as the convolution kernel. Unlike classical convolution, the convolution kernel of the dynamic hysteresis operator only needs to be dynamically changed and does not require parameter learning.
[0017] 2. A hierarchical dynamic convolutional neural network is used to model non-smooth hysteresis characteristics. In the modeling process, only the parameters of the dynamic RBF neural network need to be learned, which cleverly avoids the non-differentiability problem that exists in directly modeling hysteresis and simplifies the model complexity.
[0018] 3. The hysteresis error between the motor and reducer module is predicted by a dynamic convolutional neural network hysteresis model, and the hysteresis error is converted into the set rotation angle of the servo control motor at the module input. An open-loop feedforward compensation method is adopted to indirectly suppress and eliminate the hysteresis error, thereby effectively eliminating the hysteresis error without affecting the stability of the module system. Attached Figure Description
[0019] Figure 1 This is a characteristic diagram of the convolution hysteresis operator.
[0020] Figure 2 This is a schematic diagram of hysteresis feature pattern extraction for the convolution hysteresis operator.
[0021] Figure 3 This is a schematic diagram of the hysteresis model of a dynamic convolutional neural network.
[0022] Figure 4 This is a schematic diagram illustrating the hysteresis error compensation process of the motor and reducer module. Detailed Implementation
[0023] 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.
[0024] Unlike classical convolution, the convolution hysteresis operator (dynamic convolution) proposed in this invention uses... Constant load torque As the convolution kernel, with Time and Constant load torque The sequence being scanned is mathematically described as follows:
[0025]
[0026] In the formula, The output of the convolution hysteresis operator Time-based dynamic convolution, Input to the convolution hysteresis operator Constant load torque, Input to the convolution hysteresis operator Constant load torque.
[0027] The properties of the convolution hysteresis operator are as follows: Figure 1 As shown, it includes:
[0028] Zero return characteristics:
[0029] Returning to zero after the forward and reverse transformation from zero. ,so That is, the convolution hysteresis operator has the property of returning to zero.
[0030] Multi-value correspondence characteristics:
[0031] During the forward and reverse stroke changes, let's assume... and The amplitudes of the input signal during forward and reverse strokes are the same at both times. When, the output signals are respectively denoted as and ,but:
[0032]
[0033]
[0034] Because of the correct process Time and reverse journey These correspond to the input load torque, and ,so That is, the convolution hysteresis operator has the property of multi-valued correspondence.
[0035] It can be seen that the convolution hysteresis operator with abrupt slope changes before and after the extreme point has both zero-return property and multi-valued correspondence property, and therefore has hysteresis property.
[0036] Based on the characteristics of the convolutional hysteresis operator, the convolutional hysteresis operator can extract hysteresis feature patterns through dynamic convolution. The extracted hysteresis feature patterns describe the non-smooth hysteresis characteristics. For example... Figure 2 As shown, when the input to the convolutional hysteresis operator is a load torque time series At that time, the output is the corresponding dynamic convolution time series. The details are as follows:
[0037] It is composed of load torque in 1×2 dimensions. , ] and a 2×1 dimension load torque matrix The result of dynamic convolution;
[0038] It is a load torque in 1×2 dimensions With a 2×1 dimension load torque matrix The result of dynamic convolution;
[0039] ...;
[0040] It is a load torque in 1×2 dimensions With a 2×1 dimension load torque matrix The result is obtained by performing dynamic convolution.
[0041] in For the number of seconds.
[0042] The dynamic convolutional neural network hysteresis model proposed in this invention By convolution hysteresis operator and dynamic RBF (Radial Basis Function) neural networks Series connection, such as Figure 3 As shown. Convolution hysteresis operator. Describing strong nonlinear hysteresis feature patterns, convolutional hysteresis operator The output is obtained through a dynamic RBF neural network. The nonlinear mapping is used to obtain the hysteresis model. ,in .because It is a hierarchical structure, using a hierarchical dynamic convolutional neural network for hysteresis modeling. It only requires implementing a dynamic RBF neural network. The parameters are learned, simplifying the model complexity. Furthermore, parameter learning based on the dynamic convolutional neural network hysteresis model only involves obtaining the dynamic RBF recurrent neural network, cleverly avoiding the non-differentiability (non-differentiable) problem inherent in directly modeling hysteresis. The input to the hysteresis model is the load torque borne by the motor and reducer module. The output is the predicted torsion angle of the motor and reducer module. Based on the actual torsion angle output by the motor and reducer module. Predicted torsion angle compared with the output of the hysteresis model Based on the difference, a dynamic RBF recurrent neural network is implemented using the gradient optimization method. The parameters are learned, thereby completing the construction of the dynamic convolutional neural network hysteresis model.
[0043] Assume the output of the motor and reducer module The expected output angle at any given time is For a reduction ratio of The reducer, the motor and the reducer module input The motor rotation angle is set at all times. Ideally, the output of the motor and reducer module should be... The actual output angle at any given time is equal to The desired output angle is always present, at which point the corresponding torsional angle (hysteresis error) is zero. However, in reality, under load, the output angle of the motor and reducer module... The actual output angle at any given time is less than The desired output angle is constantly being considered, at which point the corresponding torsional angle (hysteresis error) is not zero. As the load on the motor and reducer module changes, the dynamic convolutional neural network hysteresis model adjusts accordingly. Constant load torque Predict one step ahead Predict the torsion angle at all times Based on the reduction ratio Predict the torsion angle at all times Converted to compensating rotation angle And compensate it for the motor and reducer module. The motor rotation angle is set at all times. In the middle, after compensation, the compensated rotation angle of the servo-controlled motor is obtained as follows: When the motor executes the compensated motor rotation angle value, the hysteresis error of the joint module can be effectively offset. The compensation process is as follows: Figure 4 As shown.
[0044] The method for dynamic convolutional neural network hysteresis modeling and compensation of motor and reducer modules proposed in this invention includes the following steps:
[0045] Step 1: Construct a dynamic convolutional neural network hysteresis model. The dynamic convolutional neural network hysteresis model consists of a convolutional hysteresis operator and a dynamic RBF neural network. The input of the convolutional hysteresis operator forms the input of the dynamic convolutional neural network hysteresis model, the output of the convolutional hysteresis operator is connected to the input of the dynamic RBF neural network, and the output of the dynamic RBF neural network forms the output of the dynamic convolutional neural network hysteresis model.
[0046] Step 2: Connect the motor and the reducer module. Constant load torque The result is obtained by feeding the data into a dynamic convolutional neural network hysteresis model. Predict the torsion angle at all times .
[0047] Step 3, utilize Predict the torsion angle at all times For motor and reducer modules The motor rotation angle is set at all times. To receive compensation Motor rotation angle after time compensation ,in
[0048]
[0049] In the formula, This refers to the reduction ratio between the motor and the reducer module.
[0050] In summary, this invention proposes a simple dynamic convolution hysteresis operator, and combines it with a dynamic RBF (Radial Basis Function) that can realize strong nonlinear characteristic mapping to construct a hysteresis model of the motor and harmonic reducer module. Through feedforward compensation control, the module hysteresis error is compensated, thereby improving the accuracy of the joint module.
[0051] 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 dynamic convolutional neural network hysteresis modeling and compensation method for motor and reducer modules, characterized by: Step 1: Construct a dynamic convolutional neural network hysteresis model. This dynamic convolutional neural network hysteresis model consists of a convolutional hysteresis operator and a dynamic RBF neural network. The input of the convolutional hysteresis operator forms the input of the dynamic convolutional neural network hysteresis model, the output of the convolutional hysteresis operator is connected to the input of the dynamic RBF neural network, and the output of the dynamic RBF neural network forms the output of the dynamic convolutional neural network hysteresis model. Step 2: Connect the motor and the reducer module. Constant load torque The result is obtained by feeding the data into a dynamic convolutional neural network hysteresis model. Predict the torsion angle at all times ; Step 3, utilize Predict the torsion angle at all times For motor and reducer modules The motor rotation angle is set at all times. To receive compensation Motor rotation angle after time compensation ,in: In the formula, This refers to the reduction ratio between the motor and the reducer module.
2. The method for dynamic convolutional neural network hysteresis modeling and compensation of motor and reducer modules according to claim 1, characterized in that, In step 1, the mathematical description of the convolution hysteresis operator is as follows: In the formula, The output of the convolution hysteresis operator Time-based dynamic convolution, Input to the convolution hysteresis operator Constant load torque, Input to the convolution hysteresis operator Constant load torque.