A servo motor system control parameter intelligent optimization method

Abstract

The application discloses a kind of servo motor system control parameter intelligent optimization method, this method is by constructing radial basis function optimization network, with the characteristic parameter of given position trajectory as the input of optimization network, control parameter as output, directly to servo motor system multiple control parameters are optimized;Optimization process needs to train radial basis function network, with the mean square error of given position trajectory and actual position trajectory as loss function, the undetermined parameter of radial basis function network is updated using gradient direction propagation algorithm, when network parameter converges, the output control parameter is the optimal control parameter of optimization completion.Therefore, the feature of the present application is that network structure is simple, and the amount of calculation is small, without understanding servo system control structure, can efficiently complete the simultaneous optimization of multiple parameters.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of servo motor control, and particularly relates to a servo motor system control parameter intelligent optimization method. BACKGROUND

[0002] In the field of high-precision machining, the servo motor control system is usually based on field-oriented control, and a three-loop control scheme based on a PI controller is used as a basic framework. The three-loop structure includes position control, speed control and current control, and the ultimate goal is to track the given position trajectory with high precision.

[0003] The control parameters of the three-loop structure are usually not universal, and in the face of actual complex operating conditions, the controller parameters need to be specifically set to meet the actual application requirements. The classical empirical formula setting method relies on the low-order approximate model of the motor system and the selection of some empirical coefficients, and the setting results are difficult to be used for actual high-order nonlinear systems; in actual application, trial and error setting is required with artificial experience and professional knowledge, but the parameters of the motor system are coupled with each other, and the selection of optimal parameters is difficult and time-consuming. In order to reduce the cost of professional labor and improve the setting efficiency, some automatic optimization methods are usually considered to assist parameter optimization, which can be mainly divided into: adaptive control, heuristic optimization and neural network optimization algorithm; adaptive control algorithm such as document [Xie B, Zhou LM, Niu K, et al. Research on fuzzy adaptive PID control system of wheat seeding machine seeding depth [J / OL]. Transactions of the Chinese Agricultural Machinery Society, 1-12, 2024] can automatically adjust the parameters online according to the system state change, but it may face slow convergence, high algorithm complexity and poor robustness when facing complex nonlinear systems; heuristic optimization algorithm such as document [Gaing ZL. A particle swarm optimization approach for optimum design of PID controller in AVR system [J]. IEEE Transactions on Energy Conversion, 2004, 19(2): 384-391. DOI: 10.1109 / TEC.2003.821821] uses genetic algorithm, particle swarm optimization algorithm and other methods to optimize the motor control parameters, which can effectively avoid local optimal solution, but when facing multi-coupling parameter optimization target, the hardware resource occupation of heuristic optimization is high, and the convergence efficiency is low; neural network optimization algorithm such as document [Zhang MJ, Zhang Hg. Neural network PID controller based on genetic algorithm optimization [J]. Journal of Jilin University (Engineering Version), 2005, (01): 91-96. DOI: 10.13229 / j.cnki.jdxbgxb.200501019] uses multilayer perception to fit the ability of the controller, which can quickly respond to system changes and has high optimization efficiency, but it needs to be combined with the controller structure, and when the optimization target is too much, it cannot effectively perform global search.

[0004] Therefore, the current control parameter optimization algorithm still has problems such as low efficiency, complex model and unstable optimization results when dealing with multi-parameter optimization requirements; how to design an efficient and simple control parameter automatic optimization method to quickly obtain the optimal servo motor system control parameters in specific application background is still quite challenging. SUMMARY

[0005] In view of the above, the application provides a servo motor system control parameter intelligent optimization method, which has simple network structure, small calculation amount, does not need to understand the servo system control structure, and can efficiently complete the simultaneous optimization of multiple parameters.

[0006] A servo motor system control parameter intelligent optimization method, comprising the following steps:

[0007] (1) determining the input and output of the radial basis function network;

[0008] (2) designing the structure of the radial basis function network;

[0009] (3) training the radial basis function network, and using the radial basis function network to predict and optimize the servo motor system control parameters.

[0010] Further, in the step (1), the position trajectory curve feature vector is generated according to the given reference position trajectory as the input of the radial basis function network, and the position trajectory curve feature vector is composed of four features of the motor rotor movement time length, maximum speed, maximum acceleration and maximum jerk.

[0011] In the conventional neural network optimization architecture, the network relies on dynamic information such as error as the input signal, which requires a large amount of data to achieve effective training; especially when dealing with complex situations involving multiple links and multiple parameters, the optimization process of the model often presents instability, which makes it difficult to achieve the expected training effect. In comparison, the feature vector of the position trajectory curve is extracted from the given reference trajectory to represent the parameter set of the curve feature, which remains constant during the motor control process; therefore, only the undetermined parameters of the network need to be adjusted according to the response characteristics of the motor system to optimize the motor control parameters, thereby ensuring the stability in the multi-link and multi-parameter optimization process. In addition, the dimension of the position trajectory curve feature vector is much lower than the actual point number of the trajectory curve, which significantly reduces the burden of network input data.

[0012] Further, in the step (1), the output of the radial basis function network is the servo motor system control parameter, and for the servo motor system adopting a three-loop (position loop, speed loop and current loop) control structure, the servo motor system control parameter includes position control gain, speed feedforward, speed control gain, speed control integral coefficient, acceleration feedforward, position command low-pass filter time constant, speed feedforward low-pass filter time constant, speed command low-pass filter time constant, speed feedback command low-pass filter time constant and current command low-pass filter time constant.

[0013] Further, the radial basis function network in step (2) is composed of an input layer, a hidden layer and an output layer connection, the input layer is composed of 4 neurons, which respectively correspond to the four features in the position trajectory curve feature vector; the hidden layer is composed of multiple neurons, which use radial basis functions to perform nonlinear feature mapping on the input layer neurons; the output layer is composed of multiple neurons, which are used to linearly combine the hidden layer neurons to obtain the servo motor system control parameters.

[0014] Further, the radial basis function adopts a Gaussian function, which has the following expression:

[0015]

[0016] wherein b j is the jth neuron value of the hidden layer, x pos is the position trajectory curve feature vector, c j is the jth row vector in the center point parameter matrix, σ j is the jth element value in the width parameter vector, i.e., the width parameter of the jth neuron in the hidden layer, exp() represents the natural exponential function, j is a natural number and 1≤j≤M, and M is the number of neurons in the hidden layer.

[0017] The conventional neural network optimization framework usually adopts a multi-layer perceptron network structure composed of multiple linear transformations and nonlinear activation functions. When dealing with complex optimization problems, more layers and neurons need to be configured, which increases the difficulty of network training and the complexity of hyperparameter selection. In contrast, the Gaussian radial basis function effectively captures the nonlinear relationship in the data by mapping the input curve feature vector into a high-dimensional space. After the radial basis function, the network does not need to explicitly calculate the high-dimensional feature vector, but realizes the mapping through the calculation of the inner product in the high-dimensional space. Therefore, the network based on the radial basis function can more efficiently complete the optimization task of multiple control parameters under multiple links while keeping a small scale.

[0018] Further, the expression of the output layer linearly combining the hidden layer neurons is as follows:

[0019]

[0020] wherein w ij is the element value in the i th row and j th column of the weight parameter matrix, x para,i is the i th neuron value of the output layer, i.e., the i th servo motor system control parameter, i is a natural number and 1≤i≤k, k is the number of neurons in the output layer, i.e., the number of servo motor system control parameters.

[0021] Further, the step (3) takes the way of training and predicting at the same time for the radial basis function network, and the position trajectory curve feature vector generated by the reference position trajectory is input into the radial basis function network in each control cycle, and since the radial basis function network parameters are also updated once in each control cycle, the servo motor system control parameters output by the radial basis function network are constantly optimized.

[0022] Further, the specific process of training the radial basis function network in the step (3) is as follows:

[0023] 3.1 initialize the radial basis function network parameters;

[0024] 3.2 input the position trajectory curve feature vector generated by the reference position trajectory into the radial basis function network to obtain the servo motor system control parameters;

[0025] 3.3 control the servo motor to track the reference position trajectory under the servo motor system control parameters to obtain the actual position trajectory of the servo motor;

[0026] 3.4 calculate the loss function of the radial basis function network according to the error between the actual position trajectory and the reference position trajectory;

[0027] 3.5 use the loss function to perform back propagation on the radial basis function network by gradient descent method, and update the radial basis function network parameters;

[0028] 3.6 repeatedly execute steps 3.2-3.5 to constantly update the radial basis function network parameters until the loss function or the maximum iteration number is reached.

[0029] Further, the expression of the loss function is as follows:

[0030]

[0031] Wherein: Loss represents the loss function, p ref and p real are the reference position trajectory and the actual position trajectory respectively, and T represents transposition.

[0032] Further, the radial basis function network parameters include the center point parameter matrix (M×4 dimension), the width parameter vector (M×1 dimension) and the weight parameter matrix (M×k dimension), and the iterative update formula thereof is as follows:

[0033]

[0034]

[0035]

[0036] w' = w + η1Δw / L

[0037] c' = c + η2Δc / L

[0038] σ' = σ + η3Δσ / L

[0039] wherein w and w' are weight parameter matrices before and after updating respectively, Δw is the error gradient of the weight parameter matrix, c and c' are center point parameter matrices before and after updating respectively, Δc is the error gradient of the center point parameter matrix, σ and σ' are width parameter vectors before and after updating respectively, Δσ is the error gradient of the width parameter vector, x pos is a position trajectory curve feature vector, x para is a servo motor system control parameter vector, L is the total number of reference position trajectories, η1-η3 are given learning rate coefficients, and b represents a hidden layer.

[0040] Based on the above technical solutions, the present application has the following beneficial technical effects:

[0041] 1. The parameter optimization method based on the radial basis function network has the advantages of simple structure, less undetermined parameters of the optimized network, high calculation efficiency, easy deployment on various low-computing-power control platforms, and realization of online optimization of servo motor system control parameters.

[0042] 2. The optimization method does not need to consider the internal details of the control structure and does not depend on expert knowledge in the field of servo motor control, thereby reducing the human burden.

[0043] 2. The radial basis function network can be regarded as a black box model, and only the position error is needed to simultaneously optimize multiple control parameters of the servo motor system, and the optimization process is fast and efficient. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 Fig. 1 is a structural schematic diagram of a servo motor control system in an embodiment of the present application.

[0045] Figure 2 Fig. 2 is a schematic diagram of a given reference position trajectory curve in an embodiment of the present application.

[0046] Figure 3 Fig. 3 is a schematic diagram of a radial basis function network parameter optimization process in an embodiment of the present application.

[0047] Figure 4 Fig. 4 is a schematic diagram of a loss iteration process curve of network training in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to more specifically describe the present application, the technical solutions of the present application are described in detail below in combination with the drawings and specific embodiments.

[0049] This implementation takes a small servo motor with a rated power of 0.4kW, a rated torque of 1.27Nm, and a rated speed of 3000r / min as an example to optimize multiple control parameters in the control structure.

[0050] like Figure 1 As shown, the servo motor control system adopts a three-loop control framework based on a field-oriented control strategy, which generally includes a position loop, a speed loop, and a current loop control structure. Given a reference position trajectory... L represents the total number of points on the trajectory; in this embodiment, L = 8000. Under appropriate control parameters, the actual position trajectory of the controlled motor is... trajectory with a given reference position Consistent. In the position loop structure, the given position trajectory first passes through the position command low-pass filter stage C1 to obtain a relatively smooth position command θ0. Then, θ0 is compared with the actual trajectory feedback, and after passing through the position scaling stage C2, a velocity command s1 is obtained to improve the steady-state accuracy of the system. At the same time, the velocity feedforward filter stage C3 and the velocity feedforward stage C4 generate a direct velocity command s2 to improve the dynamic response of the system and improve tracking accuracy. The sum of s1 and s2 is used as the velocity command s. ref It acts on the subsequent velocity loop. In the velocity loop structure, given s ref After passing through a speed command low-pass filter C5, a smooth speed command s0 is obtained. This is then compared with the speed feedback, and after passing through a speed proportional-integral filter C6, the q-axis current command i is obtained. q1 Simultaneously, given s0, a direct current i is generated through the acceleration feedforward stage C7. q2 This is used to improve dynamic performance and enhance system stability; the velocity feedback is derived from the position feedback, requiring the addition of a velocity feedback low-pass filter C8 to smooth the feedback velocity signal. q1 and i q2 The superposition generates the q-axis current given i. qref This signal is then applied to the subsequent q-axis current loop. After passing through the current command low-pass filter stages C9 and C10, the current loop is compared with the current feedback. After passing through the proportional-integral stages C11 and C12, the motor voltage vector is obtained, which then controls the motor to run along the desired trajectory. Furthermore, to improve system stability, anti-interference capability, and reduce the impact of noise on control accuracy, a series of filters are incorporated into the three-loop control structure of the servo motor control system. The filter parameters may affect the system's response speed; therefore, a trade-off must be struck between noise suppression and response speed to ensure overall system performance.

[0051] To achieve the goal of optimizing control parameters, this embodiment establishes an optimization network with radial basis functions as its core, given the ground position trajectory curve p. refThe input is input to the optimization network, and the output is the control parameter to be optimized, and then these parameters are used to control the motor to track the given trajectory with high precision. The initial network parameters cannot output the optimal control parameters, and the network parameters need to be adjusted according to the given reference position trajectory curve and the current tracking error for multiple rounds until the optimal control parameters are output. The specific implementation process is as follows:

[0052] (1) Determine the input of the radial basis function network.

[0053] In order to unify the input of the network and reduce the size of the network, the parameters are generated according to the curve of the given reference position trajectory. In this embodiment, the reference position trajectory adopts an S-shaped trajectory curve, which can smoothly control the change of speed and acceleration, reduce the impact and vibration of the system, and improve the stability and precision of the motion; The trajectory is usually divided into three stages, namely acceleration stage, constant speed stage and deceleration stage, which can be described by four characteristic parameters of the motion time length of the motor rotor, maximum speed, maximum acceleration and maximum jerk. Therefore, the four characteristic parameters can be combined into a feature vector as the input of the network. In this embodiment, the motion time length of the motor rotor is set to 2s, the maximum speed is 8r / s, the maximum acceleration is 2000r / s 2 , and the maximum jerk is 1000r / s. The reference position trajectory curve is shown in Figure 2 .

[0054] (2) Determine the output of the radial basis function network.

[0055] The output of the network is the k control parameters to be optimized. The selection of the optimization parameters needs to consider the structure of the three-loop control system set. The current loop is the basic link in the control structure, which is responsible for adjusting the current to achieve the required torque output, but it mainly affects the fast response and anti-interference ability of the system; In the trajectory tracking scene, the direct influence of the current loop on the trajectory tracking is relatively small, and more attention is usually paid to the position loop control parameters and the speed loop control parameters.

[0056] In this embodiment, the parameters to be optimized include the position control gain K pp , the speed feedforward K vf , the speed control gain K vp , the speed control integral coefficient K vi , the acceleration feedforward K af , the position command low-pass filter time constant T p , the speed feedforward low-pass filter time constant T pvf , the speed command low-pass filter time constant T v , the speed feedback command low-pass filter time constant T vf , and the current command low-pass filter time constant T c, i.e. the radial basis function network has 10 outputs, and can optimize 10 key control parameters simultaneously.

[0057] (3) Determine the radial basis function network structure.

[0058] The radial basis function network in this embodiment has three layers, i.e. an input layer, a hidden layer and an output layer. The input layer transmits the original input to the hidden layer, which is composed of d neurons and corresponds to x pos ; the output layer is a linear combination of the results of the hidden layer, and generates the optimized parameters, which is composed of k neurons and corresponds to x para ; the hidden layer performs nonlinear feature mapping, which greatly affects the final output result, and is composed of M neurons. In this embodiment, d = 4, k = 10, and the number of hidden layer neurons M is set to 20 considering the high-dimensional feature processing capability and computational complexity of the network.

[0059] The input x pos is obtained through the radial basis function activation function, and then linearly combined to obtain k optimized parameters; the radial basis function adopts a Gaussian function, and its expression is:

[0060]

[0061] In the formula, b j is the jth neuron element in the hidden layer ; c is the jth row vector in the center point parameter matrix , which represents the center point of each j hidden layer neuron; and σ j is the jth element in the width parameter vector , which represents the width parameter of each j neuron.

[0062] The output x para of the network is a linear combination of the sum of the basis functions and the connection weights, and its form is:

[0063]

[0064] In the formula, w ij is the i-row j-column element in the weight parameter matrix , which represents the linear combination weight.

[0065] (4) Network training and control parameter optimization.

[0066] The weight parameters w, center point parameters c and width parameters σ of the radial basis function network are to be determined, and need to be trained through data. After obtaining the stable converged network parameters, the network output is the optimized control parameters. The radial basis function network optimization process is as follows: Figure 3As shown, the corresponding network updating process can include the following detailed steps:

[0067] Step 1: initialize the radial basis function network parameters and control parameters x para .

[0068] Step 2: under the current control parameters x para , control the motor to track the given reference position trajectory p ref , and obtain the actual position trajectory p real .

[0069] Step 3: use the error between the reference position trajectory p ref and the actual position trajectory p real to calculate the loss function of the radial basis function network, which adopts the least mean square error, and its form satisfies:

[0070]

[0071] Step 4: use the back propagation algorithm to calculate the gradient of the error and propagate backward layer by layer, and update the network parameters in the average gradient direction, and the update formula of the weight parameter w, the center point parameter c and the width parameter σ is:

[0072]

[0073]

[0074]

[0075] w' = w + η1Δw / L

[0076] c' = c + η2Δc / L

[0077] σ' = σ + η3Δσ / L

[0078] The radial basis function network is applied to optimize the control parameters, and the loss change curve in the iteration process is as shown in Figure 4 After 60 rounds of training, the parameters to be optimized have converged to a stable value. It should be noted that the read-back position is a discrete integer quantity, Figure 4 the initial loss is 4.30 x 10 9 , and the loss value at the end of training is 1.67 x 10 7 . As can be seen from the loss change, after the radial basis function network optimization of the servo motor system control parameters, the position tracking error is significantly reduced, by about 258 times.

[0079] The following table compares the position tracking error results of the servo motor system under the optimized parameters obtained by the embodiment, the initial training parameters and the manually adjusted parameters.

[0080] Table 1

[0081]

[0082] The ISE represents the sum of squares of the deviation between the given position trajectory and the actual position trajectory. The results show that the ISE index after using the optimization parameters obtained by the embodiment is significantly reduced, which is reduced by 867 times compared with the manual adjustment parameters and reduced by 3933 times compared with the initial training parameters. Obviously, after optimization, the position tracking accuracy is significantly improved.

[0083] The above description of the embodiments is to facilitate those of ordinary skill in the art to understand and apply the present application. Those skilled in the art can easily make various modifications to the above embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present application is not limited to the above embodiments, and any improvements and modifications made to the present application by those skilled in the art based on the disclosure of the present application shall be within the scope of protection of the present application.

Claims

1. A method for intelligent optimization of control parameters of a servo motor system, comprising the following steps: (1) Determine the input and output of the radial basis function network, wherein the input of the radial basis function network includes a given reference position trajectory curve, and the output of the radial basis function network includes multiple control parameters to be optimized, which are associated with the structure of the three-loop control system; (2) Design the structure of the radial basis function network; (3) Train the radial basis function network and use the radial basis function network to predict and optimize the control parameters of the servo motor system; The specific process of training the radial basis function network in step (3) is as follows: 3.1 Initialize the radial basis function network parameters; 3.2 The feature vector of the position trajectory curve generated from the reference position trajectory is input into the radial basis function network to obtain the control parameters of the servo motor system; 3.3 Under the control parameters of the servo motor system, control the servo motor to track the reference position trajectory and obtain the actual position trajectory of the servo motor; 3.4 Calculate the loss function of the radial basis function network based on the error between the actual position trajectory and the reference position trajectory; 3.5 Backpropagation of the radial basis function network is performed using gradient descent with the loss function, and the parameters of the radial basis function network are updated. 3.6 Repeat steps 3.2 to 3.5 to continuously update the radial basis function network parameters until the loss function or the maximum number of iterations is reached; The step of predicting and optimizing the control parameters includes: adjusting the network parameters in multiple rounds based on the reference position trajectory curve and the current tracking error until the optimal control parameters are output.

2. The intelligent optimization method for control parameters of a servo motor system according to claim 1, characterized in that: In step (1), a position trajectory curve feature vector is generated based on the given reference position trajectory as the input of the radial basis function network. The position trajectory curve feature vector is composed of four features: the motion time length of the motor rotor, the maximum speed, the maximum acceleration, and the maximum jerk.

3. The intelligent optimization method for control parameters of a servo motor system according to claim 1, characterized in that: The output of the radial basis function network in step (1) is the control parameters of the servo motor system. The servo motor system adopts a three-loop control structure. The control parameters of the servo motor system include position control gain, speed feedforward, speed control gain, speed control integral coefficient, acceleration feedforward, position command low-pass filter time constant, speed feedforward low-pass filter time constant, speed command low-pass filter time constant, speed feedback command low-pass filter time constant, and current command low-pass filter time constant.

4. The intelligent optimization method for control parameters of a servo motor system according to claim 1, characterized in that: In step (2), the radial basis function network is composed of an input layer, a hidden layer and an output layer. The input layer consists of 4 neurons, which correspond to the four features in the feature vector of the position trajectory curve. The hidden layer consists of multiple neurons, which use radial basis functions to perform nonlinear feature mapping on the input layer neurons; the output layer consists of multiple neurons, which are used to linearly combine the hidden layer neurons to obtain the control parameters of the servo motor system.

5. The intelligent optimization method for control parameters of a servo motor system according to claim 4, characterized in that... It lies in: The radial basis function is a Gaussian function, and its expression is as follows: ； in: Let j be the value of the j-th neuron in the hidden layer. The feature vector of the position trajectory curve. Let j be the vector in the j-th row of the center point parameter matrix. exp() represents the width parameter of the j-th element in the width parameter vector, which is the width parameter of the j-th neuron in the hidden layer. j is a natural number and 1≤j≤M, where M is the number of neurons in the hidden layer.

6. The intelligent optimization method for control parameters of a servo motor system according to claim 5, characterized in that: The expression for the linear combination of hidden layer neurons in the output layer is as follows: ； in: The value of the element in the i-th row and j-th column of the weight parameter matrix. Let i be the value of the i-th neuron in the output layer, which is the i-th servo motor system control parameter. i is a natural number and 1≤i≤k, where k is the number of neurons in the output layer, which is the number of servo motor system control parameters.

7. The intelligent optimization method for control parameters of a servo motor system according to claim 1, characterized in that: In step (3), the radial basis function network is trained and predicted simultaneously. In each control cycle, the feature vector of the position trajectory curve generated by the reference position trajectory is input to the radial basis function network. Since the radial basis function network parameters are also iteratively updated once in each control cycle, the servo motor system control parameters output by the radial basis function network are continuously optimized each time.

8. The intelligent optimization method for control parameters of a servo motor system according to claim 1, characterized in that: The expression for the loss function is as follows: ； Where: Loss represents the loss function. and These are the reference position trajectory and the actual position trajectory, respectively, with T representing transpose.

9. The intelligent optimization method for control parameters of a servo motor system according to claim 7, characterized in that: The radial basis function network parameters include a center point parameter matrix, a width parameter vector, and a weight parameter matrix, and their iterative update formula is as follows: ； ； w'=w+ △w / L; c'=c+ △c / L; σ'=σ+ △o / L; Where: w and w' are the weight parameter matrices before and after the update, respectively; Δw is the error gradient of the weight parameter matrix; c and c' are the center point parameter matrices before and after the update, respectively; Δc is the error gradient of the center point parameter matrix; σ and σ' are the width parameter vectors before and after the update, respectively; Δσ is the error gradient of the width parameter vector. and These are the reference position trajectory and the actual position trajectory, respectively. The feature vector of the position trajectory curve. Here, L represents the control parameter vector for the servo motor system, and L represents the total number of points on the reference position trajectory. ~ Given the learning rate coefficient, b represents the hidden layer.

Images