A method for realizing a neural network compensation motor load torque observer

Abstract

The present invention relates to a method for realizing a neural network compensated motor load torque observer, and belongs to the field of motor control. The load torque observer of the present invention is based on a motor kinematic model, and a neural network is added to perform uncertainty compensation on the basis of analyzing the uncertainty of its parameters. The designed torque observer takes the motor q-axis current and the motor speed as input, and outputs the motor load observation torque. The neural network uses a radial basis neural network, which continuously approaches the model error through online training, selects a suitable online weight adaptive update law, and decides whether to add neural network compensation according to the error, so as to realize the switching between an ordinary observer and an observer with a neural network. The method designed by the present invention can ensure the stability and rapid adaptability of the observer, and has higher accuracy, a simple system structure, and is easier to be applied in engineering practice.

Description

Technical Field

[0001] The invention belongs to the field of motor control. Background Art

[0002] With the continuous development of the national economy, production efficiency has gradually improved, and product quality has become increasingly higher. As the core component of the entire industrial automation field, the motor has higher and higher performance requirements. Permanent magnet synchronous servo motors are widely used in the field of high-precision servo because of their high power density ratio, low cogging torque, high working efficiency and low power consumption. Research on its related control technology has important practical significance. High-performance applications require that the output of the servo motor has no overshoot and quickly tracks the input command, and has no static error in the steady state, and has strong robustness to changes in system parameters. At present, in the actual control of motors, the most commonly used is speed and torque loop PI control. When there is a disturbance in the load or a sudden change in torque, the robustness and anti-interference of the system are not strong enough, and the speed will fluctuate significantly, which seriously affects the performance of the system. When the load changes frequently or the disturbance torque persists without any adjustment to the original control system, the stability and robustness of the entire system will be greatly affected. Summary of the invention

[0003] In order to solve the problems of load disturbance and mutation that are difficult to overcome in the dual closed-loop control method of motor speed loop and current loop, the present invention provides a method for realizing a neural network online compensation motor load torque observer, which can estimate the load torque of the motor during the motor control process and then add it to the controller to improve the stability and robustness of the motor control.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for realizing a neural network online compensation motor load torque observer comprises the following steps:

[0006] Step 1: Establish the ideal kinematic model of the motor;

[0007] Step 2: Since the motor parameters are unknown and not constant when the motor is modeled, there is uncertainty. Analyze the uncertainty of the motor model.

[0008] Step 3: Design and analyze the observer, and design a radial basis neural network to compensate for the uncertainty of the observer;

[0009] Step 4: Design an adaptive law for the weights of the radial basis neural network. The network weights are adjusted periodically according to the adaptive law to continuously approach the optimal output of the network.

[0010] Step 5: Determine whether to add neural network compensation based on the error of the motor speed feedback state quantity, set the error threshold, add neural network compensation when the error is greater than or equal to the threshold, and exit the neural network compensation when the error is less than the threshold and lasts for a period of time (to prevent oscillation caused by frequent switching).

[0011] Furthermore, the step 3 adopts a dimensionality reduction observer, the dimensionality reduction observer of the motor load torque takes the q-axis current and the motor speed of the motor as input, and outputs the observed value of the motor load torque. The neural network adopts a radial basis neural network, takes the state quantity error as the network input, and outputs the unknown function of the model. Approaching uncertainty.

[0012] By adopting the above technical approach, the present invention has the following beneficial effects:

[0013] The method for realizing a neural network online compensation motor load torque observer of the present invention utilizes a radial basis neural network to compensate for the uncertainty of a motor model, and the motor load torque observer designed on this basis has higher accuracy and reliability than an observer designed solely relying on a motor model. The observation result of the load torque is used for motor motion control, which can achieve better control of the motor. The selected neural network has a single hidden layer, a simple structure, and is easy to apply in engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Design a flow chart for the torque observer;

[0015] Figure 2 This is the block diagram of the torque observer system. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be described in detail below in conjunction with specific implementation methods and the accompanying drawings of the present invention.

[0017] The design flow chart of the load torque observer involved in the present invention is as follows Figure 1 As shown, the system structure diagram involved is as follows Figure 2 As shown. Step 1 is to establish the ideal kinematic model of the motor. Step 2 is to analyze the uncertainty of the motor model caused by the unknown and non-constant motor parameters when the motor is modeled. Step 3 is to design and analyze the dimension reduction observer, and design the uncertainty part of the radial basis neural network compensation observer. Step 4 is to design the adaptive law of the radial basis neural network weights to continuously approach the optimal output of the network. Step 5 is to set the error threshold. When the error is greater than or equal to the threshold, the neural network compensation is added. When the error is less than the threshold and lasts for a period of time, the neural network exits the compensation.

[0018] The motor kinematic model in step 1 is as follows:

[0019]

[0020] Where: J is the motor moment of inertia;

[0021] ω m is the motor rotor mechanical angular velocity;

[0022] B m is the viscous damping coefficient;

[0023] T L is the load torque;

[0024] T e is the motor output torque;.

[0025] For the analysis of model uncertainty in step 2, the model is expressed as a state space equation:

[0026]

[0027] in: C = [1 0]; u = T e ; y = ω m .

[0028] Right now:

[0029]

[0030] Taking into account the inaccuracy of the parameters in the model, the state equation of the motor is re-expressed as follows:

[0031]

[0032] Among them, the function f m Satisfy the following formula:

[0033] ΔAx=Bf m (5)

[0034] Where: ΔA represents the uncertainty of model parameters.

[0035] In the above step 3), the state observer of the motor can be designed as a reduced-dimensional observer in the following form:

[0036]

[0037] Where: is the estimated value of the state variable x;

[0038] K is the gain matrix of the observer;

[0039] f m An estimated value of

[0040] is the estimated value of the output y.

[0041] For the above model-based dimension reduction observer, a radial basis neural network is designed to compensate for model uncertainty. The output of the neural network is Approximating the unknown function f m ,The input of the neural network is the error value between the measured value of the motor speed and the observed value of the speed.

[0042]

[0043] Where: is the weight of the neural network, approaching the optimal weight output θ;

[0044] h is the Gaussian basis function of the network.

[0045] Then, formula (5) can be re-expressed as follows:

[0046]

[0047] Based on the neural network designed in step 3, the following error term is defined:

[0048]

[0049] The error equation of the state observer is as follows:

[0050]

[0051] Define the Lyapunov function as: tr(.) is the trace of the matrix. The matrix P is a symmetric positive definite matrix and satisfies the equation PA+A T P=-Q,Q≥0.

[0052] Design the neural network weight update rate:

[0053]

[0054] Where: η is an adjustable parameter;

[0055] For step 5 above, if Figure 2 As shown, according to the error of the motor speed state quantity Determine whether to add neural network compensation output Set the error threshold ψ. When the error is greater than or equal to the threshold ψ, neural network compensation is added. When the error is less than the threshold and lasts for a period of time ΔT, the neural network exits compensation, that is, the observer becomes the following form:

[0056]

Claims

1. A method for realizing a neural network compensated motor load torque observer, characterized in that: Step 1: Establish the ideal kinematic model of the motor: Where: J is the motor's moment of inertia; ω m is the mechanical angular velocity of the motor rotor; B m is the viscous damping coefficient; T L is the load torque; T e is the motor output torque; Step 2: Analyze the uncertainty of the motor model caused by unknown motor load parameters and non-constant damping coefficient during motor modeling, and express the model as a state space equation: Among them: C=[1 0];u=T e ;y=ω m ; Taking into account the unknown load parameters and the non-constant damping coefficient in the model, the state equation of the motor is re-expressed as follows: Among them, the function f m Satisfy the following formula: ΔAx=Bf m Where: ΔA represents the uncertainty of model parameters; Step 3: Design and analyze the state observer, and design the uncertainty part of the radial basis neural network compensation observer; the load torque observer of the motor is designed as a reduced dimension observer in the following form: Where: is the estimated value of the state variable x; K is the gain matrix of the observer; f m An estimated value of is the estimated value of the output y; Based on the above model-based dimension reduction observer, a radial basis neural network is designed to compensate for model uncertainty. The output of the neural network is Approximating the unknown function f m ,The input of the neural network is the error value between the measured value of the motor speed and the observed value; Where: is the weight of the neural network, approaching the optimal weight output θ; h is the Gaussian basis function of the network; The dimension reduction observer can then be reformulated as follows: Step 4: Design the adaptive law of the radial basis neural network weights as The matrix P is a symmetric positive definite matrix and satisfies the equation PA+A T P = -Q, Q ≥ 0, η is an adjustable parameter, thus continuously approaching the optimal output of the neural network; Step 5: Set the error threshold. When the error is greater than or equal to the threshold, the radial basis neural network is added for compensation. When the error is less than the threshold and lasts for a period of time, the radial basis neural network exits compensation. The threshold is set according to the actual indicator accuracy requirements.

2. The method for realizing a neural network compensated motor load torque observer according to claim 1, characterized in that: The step 3 adopts a dimensionality reduction observer. The dimensionality reduction observer of the motor load torque takes the q-axis current and the motor speed of the motor as input and outputs the observed value of the motor load torque. The neural network adopts a radial basis neural network, takes the state quantity error as the network input, and outputs the unknown function of the model. Approaching uncertainty.

Images