Preset performance sliding mode control method for servo system based on spectral normalized neural network

Abstract

The present disclosure provides a servo system preset performance sliding mode control method based on a spectral normalized neural network. The method establishes a first servo system model without disturbance and a second servo system model with a wind disturbance model. Based on the same control variable input, the two models generate system state data, and the difference in the system state data is used to calculate the wind torque T. d , obtain training samples; use the training samples to train the spectral normalization neural network so that it has wind moment T d The proposed method uses a spectral normalization neural network to estimate the wind torque based on the system status data output by the turntable servo system. This wind torque estimate is then applied to a sliding mode controller with preset performance to generate the control variable u. This method can improve the robustness of the control scheme.

Description

Technical Field

[0001] The present invention relates to the field of actual servo system compensation control, and in particular to a spectral normalization neural network preset performance sliding mode control method for a servo system disturbed by gusts of wind. Background Art

[0002] With the rapid development of modern industry, key sectors such as precision manufacturing, aerospace, and robotics have placed increasingly stringent demands on the control accuracy of servo systems. The control accuracy of a servo system is not only directly related to the system's tracking performance but also has a profound impact on the stability and response speed of the entire system. However, in actual operation, servo systems are often constrained by multiple factors such as nonlinear friction and unknown external interference, making it difficult to further improve system control accuracy and seriously hindering the further development of high-end manufacturing and automation technologies.

[0003] Deep learning, particularly neural network architectures, has demonstrated remarkable achievements and broad application potential in fields such as image generation, semantic segmentation, and reinforcement learning. However, neural networks often face stability challenges during training, including critical issues such as vanishing or exploding gradients. These issues hinder model convergence and ultimately impact performance. In the specific field of generative adversarial networks (GANs), the Jensen-Shannon (JS) divergence employed in the objective function optimization process often leads to vanishing gradients in the generator, further increasing the difficulty and uncertainty of training. To address this challenge, spectral normalization (SNR) has emerged as a key innovation in addressing neural network stability. This technique calculates the spectral norm (maximum singular value) of each layer's weight matrix and divides it by the spectral norm to ensure that the output of the network layer does not fluctuate too drastically. In GANs, spectral normalization is primarily used to constrain the discriminator to satisfy the 1-Lipschitz condition. Meeting this condition effectively mitigates instabilities during training, significantly improving GAN training efficiency and model performance.

[0004] Sliding mode control is a nonlinear robust control method. Its core lies in designing a model-independent sliding surface. Feedback and robust compensation ensure that the system state first reaches the sliding surface and then slides along it to the equilibrium point. During this process, the system state motion is independent of the model and disturbances, resulting in strong robustness to uncertain disturbances. To optimize performance, strategies such as integral sliding mode to reduce error, terminal sliding mode to promote finite-time convergence, and fast terminal sliding mode to increase speed have been developed. However, to cope with unknown disturbances, the sign function terms often used cause the sliding mode variables to switch near the sliding surface, leading to chattering.

[0005] The stated-performance control strategy aims to transform the system tracking error into an unconstrained form through a performance function and, through control methods, ensure that the converted error remains continuous and well-bounded, thus meeting the stringent requirements for system tracking performance. However, this approach may face the challenge of singularity: when the tracking error approaches zero, the calculated control variable may increase abnormally, causing the system to exceed the preset performance boundary and potentially become unstable. Furthermore, in actual engineering applications, situations such as sudden switching of work tasks, rapid changes in external disturbances, or the difficulty in accurately predicting the initial tracking error may cause the system tracking error to exceed the specified performance range. Summary of the Invention

[0006] In view of this, in order to solve the control problem of the servo system under gust interference, the present invention provides a preset performance sliding mode compensation control method based on spectral normalization neural network, which improves the robustness of the control scheme.

[0007] In order to solve the above technical problems, the present invention is implemented as follows.

[0008] A servo system preset performance sliding mode control method based on spectral normalization neural network, comprising:

[0009] Step 1: For the turntable servo system, build a disturbance-free first servo system model;

[0010] Step 2: Construct a wind disturbance model and convert the wind moment T output by the wind disturbance model into d Add the first servo system model to obtain the second servo system model including gust interference;

[0011] Step 3: The first servo system model and the second servo system model each generate system state data based on the same control variable input, and the wind torque T is calculated based on the difference between the system state data generated by the two servo system models. d ; The system state data generated by the second servo system model containing gust interference and the calculated wind torque T d , forming training samples;

[0012] Step 4: Use the training samples to train the spectral normalization neural network to have wind moment T d Estimation ability;

[0013] Step 5: Construct a preset performance sliding mode controller: The tracking error e of the turntable servo system is first converted into a conversion error ε according to the preset performance function, and then the conversion error ε is used for sliding mode control to output the control variable u;

[0014] Step 6: During actual control, the wind torque estimation is obtained using the spectral normalization neural network according to the system status data output by the turntable servo system. Estimate the wind moment The compensation is fed into the preset performance sliding mode controller to generate the control variable u.

[0015] Preferably, the disturbance-free first servo system model is:

[0016]

[0017] The second servo system model with gust interference is:

[0018]

[0019] in,

[0020] q is the angular position, is the angular velocity; J is the motor inertia; K1=K T / R,K2=K T K E / R,K E =n P ψ f ; R is the rotor resistance, K T is the torque time constant, T L is the system load torque, n p is the number of electrodes, ψ f is the rotor flux; y is the angular position of the turntable servo system.

[0021] Preferably, in the wind interference model, the wind speed v is the steady-state wind speed v m and pulsating wind speed v g sum;

[0022] The steady-state wind speed Among them, v m is the steady-state wind speed at height h, v mh is the reference height h m The steady-state wind speed at , α is the ground roughness index;

[0023] The pulsating wind speed v g The method of obtaining is as follows: construct a filter function H so that its frequency response curve is close to the Davenport spectrum curve; pass the white noise with zero mean through this filter function H, and the filter function H outputs the pulsating wind speed v g ;

[0024] Substitute the wind speed v model into the following wind moment T d Calculation formula to obtain wind interference model:

[0025] T d =k t pAD

[0026] Among them, k tRepresents the wind moment coefficient, p is the wind pressure, p=1 / 2ρv 2 , ρ is the air density, A is the area of ​​the circular cross-section of the turntable servo system, and D is the diameter of the circular cross-section of the turntable servo system.

[0027] Preferably, the filter function H is:

[0028]

[0029] Where s is the Laplace operator.

[0030] Preferably, the input of the spectral normalization neural network is the system state data of the turntable servo system, including the angular position q and the angular velocity

[0031] Preferably, in step 3, the wind torque T is calculated based on the difference of the system state data generated by the two servo system models. d for;

[0032] Get the output angular velocity of the two servo system models The difference Calculating wind torque

[0033] Preferably, in step 5, the tracking error e of the turntable servo system is first converted into a conversion error ε according to a preset performance function:

[0034]

[0035] Where λ(t) = e(t) / ρ(t), e(t) is the tracking error of the turntable servo system, and ρ(t) is ρ(t) = (ρ0-ρ ∞ )e -at +ρ ∞ is the performance function, ρ0 is the initial value range of the tracking error specified by the performance function, ρ ∞ is the steady-state boundary of the tracking error when the time t specified by the performance function approaches positive infinity; a is the attenuation factor; the generated conversion error ε(t) is used as the error in the sliding surface.

[0036] Preferably, the sliding surface in the preset performance sliding mode controller is a linear sliding surface s: Where c is an adjustable gain greater than zero.

[0037] Preferably, the preset performance sliding mode controller is:

[0038]

[0039] In the above formula, q is the angular position, is the angular velocity, x 1dis the given angle, J is the motor inertia, T L is the system load torque; η and k are parameters of the sliding surface, both of which are constants greater than zero; sgn(·) is the sign function, s represents the sliding surface; K1 = K T / R,K2=K T K E / R,K T is the torque time constant, R is the rotor resistance, K E =n P ψ f , n p is the number of electrodes, ψ f is the rotor flux.

[0040] Beneficial effects:

[0041] (1) This paper designs a spectral normalization neural network to approximate wind disturbances for unknown wind gusts. This approach leverages the powerful learning capabilities of neural networks to predict and compensate for complex disturbances caused by gusts, effectively suppressing these disturbances and improving the robustness of the control scheme. Spectral normalization ensures the stability of the neural network during training, thereby improving its prediction accuracy and generalization capabilities.

[0042] (2) The present invention constructs a wind interference model through numerical simulation and generates training data for training a spectral normalization neural network to predict wind interference.

[0043] (3) In order to ensure the tracking performance, a preset performance sliding mode compensation controller is designed, which combines the prescribed performance control strategy with the sliding mode controller, and limits the position error to the performance function constraint range through the preset performance function, ensuring the convergence of the tracking error, making the control scheme of the present invention more robust and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the preset performance sliding mode compensation control method based on spectral normalization neural network of the present invention.

[0045] Figure 2 This is a control block diagram of the preset performance sliding mode compensation control method based on spectral normalization neural network of the present invention.

[0046] Figure 3 Schematic diagram of the turntable servo system structure considered in the present invention.

[0047] Figure 4 Schematic diagram of wind force and wind torque.

[0048] Figure 5 The wind speed and wind torque of the wind interference model simulated according to the selected parameters.

[0049] Figure 6The wind torque is obtained by numerical simulation.

[0050] Figure 7 This is the position tracking and position error of the system under the preset performance sliding mode compensation control based on the spectral normalized neural network proposed in the present invention.

[0051] Figure 8 This is the speed tracking and speed error of the system under the preset performance sliding mode compensation control based on the spectral normalization neural network proposed in the present invention.

[0052] Figure 9 These are the approximate value and true value of wind interference using the neural network method proposed in this invention.

[0053] Figure 10 Trajectory tracking and position error of the system under sliding mode control with preset performance without compensation.

[0054] Figure 11 The speed tracking and speed error of the system under sliding mode control are the preset performance without compensation.

[0055] Figure 12 The trajectory tracking and trajectory error of the system under sliding mode control without compensation.

[0056] Figure 13 Velocity tracking and velocity error of the system under uncompensated sliding mode control. DETAILED DESCRIPTION

[0057] The present invention provides a spectral normalized neural network preset performance sliding mode control method for a servo system subject to gust interference. In order to achieve external gust interference compensation, the method uses a spectral normalized neural network to estimate and compensate for the external gust interference in the servo system. The preset performance sliding mode controller can adjust the controller output according to the degree of proximity between the system state and the predefined constraints, thereby ensuring that the servo system can follow the reference trajectory within the specified state limits and ensuring the convergence of the tracking error.

[0058] To train the spectral normalization neural network, the present invention first performs dynamic modeling of the turntable servo system, constructing two servo system models, one without wind disturbance and one with wind disturbance. The same control variable is input to both models, and the angular position output is used to calculate the angular velocity. Based on the difference in angular velocity between the two models, wind torque sample data is calculated. The spectral normalization neural network is then trained to predict wind torque, enabling it to be compensated in the controller during actual control.

[0059] The present invention is described in detail below with reference to the accompanying drawings and embodiments. Figure 1 The flowchart of the servo system preset performance sliding mode control method based on spectral normalization neural network of the present invention is shown. Figure 2The control block diagram of the present invention is shown. As shown in the figure, the method includes the following steps:

[0060] Step 1: For the turntable servo system, build a disturbance-free first servo system model.

[0061] The present invention considers Figure 3 The turntable servo system shown in the figure includes a permanent magnet synchronous motor, an encoder, a power supply system, a velocity loop controller, and a position loop controller. The permanent magnet synchronous motor precisely controls the turntable position and speed via voltage. The encoder measures position and speed, providing closed-loop feedback. The power supply system supplies power to the system. The velocity loop controller regulates motor speed, receives encoder feedback, compares it with a setpoint, and adjusts the motor output to achieve speed control. The position loop controller ensures the motor reaches the desired position, receives encoder position feedback, compares it with a setpoint, and adjusts the motor output to achieve precise position control.

[0062] The turntable servo system model considered in the present invention is as follows:

[0063]

[0064] Among them, i d is the d-axis current, u q is the current on the q axis, u d is the voltage of the d-axis, u q is the voltage of the q axis, R is the rotor resistance, L is the rotor inductance, n p is the number of electrodes, ψ f is the rotor flux, K T is the torque time constant, J is the motor inertia, T L is the system load torque, T f is the friction torque of the system, T d is the system disturbance torque, q is the angular position, is the angular velocity.

[0065] In the actual design process, the d-axis reference current It is usually set to zero, so formula (1) can be expressed as

[0066]

[0067] where K E =n P ψ f .

[0068] In addition, in reality, the electromagnetic time constant L / R is very small compared to the mechanical time constant and can be approximated to zero. The state variables are defined as y is the output, then formula (2) can be expressed as

[0069]

[0070] Where K1 = K T / R,K2=K T K E / R.

[0071] The disturbance-free first servo system model established in this step does not include disturbance, so T d and T f Remove it, and the first servo system model constructed is:

[0072]

[0073] Step 2: Construct a wind disturbance model and convert the wind moment T output by the wind disturbance model into d The first servo system model is added to obtain the second servo system model including gust interference.

[0074] This step includes the following sub-steps:

[0075] Step 21: First, numerical simulation method is used to model wind disturbance v.

[0076] The main methods for determining wind interference models are numerical simulation, field testing, and wind tunnel experiments. This paper uses numerical simulation to construct a wind interference model. Wind can be divided into steady-state wind and pulsating wind based on its periodic characteristics. Steady-state wind has a long period, and its direction and speed do not change over time. Pulsating wind has a short period, and its direction and speed vary randomly over time.

[0077] The wind speed v can be expressed as the sum of the steady-state wind speed and the fluctuating wind speed, that is:

[0078] v=v m +v g (5)

[0079] where v m is the average steady-state wind speed, v g Pulsating wind speed.

[0080] ① Steady-state wind model v m

[0081] Steady-state wind speed decreases with decreasing altitude and is often determined using an average wind speed profile curve, which mainly includes logarithmic wind speed profile curves and exponential wind speed profile curves. The exponential wind speed profile model is widely used in engineering due to its simple calculation and high accuracy. Based on observational data, Davenport et al. derived the relationship between exponential wind profiles in different sites, namely:

[0082]

[0083] where v m is the steady-state wind speed at height h, v mhis the reference height h m The steady-state wind speed at , α is the ground roughness index, which is related to the topography. The reference values ​​of the ground roughness index for common terrains are shown in Table 1 below:

[0084] Table 1 Reference values ​​of surface roughness index for common terrains

[0085]

[0086] ② Pulsating wind model v g

[0087] Analysis of the measured samples of fluctuating wind speed shows that the fluctuating wind speed can be regarded as a Gaussian stationary random process with a mean of 0 (ignoring the average wind speed). Its power spectrum conforms to the Davenport spectrum function, so the function can be used to process the Gaussian white noise to obtain the fluctuating wind speed v g Davenport spectral function v Affected by the average wind speed and ground roughness, it can be expressed as

[0088]

[0089]

[0090] where k s is the surface tension coefficient, w is the wind frequency, z is the height from the ground, and z0 is the ground roughness height.

[0091] When simulating wind speed, first construct a filter function H so that its frequency response curve is close to the Davenport spectrum curve; then, pass the white noise with zero mean through this filter function H to obtain the pulsating wind speed v g The filter function H is constructed as follows:

[0092]

[0093] Where s is the Laplace operator.

[0094] ③ Wind torque modeling

[0095] Schematic diagram of wind force and wind torque Figure 4 As shown, based on the above steady-state wind model and pulsating wind model, and combined with formula (5), the wind speed v can be obtained, and then the wind torque T can be calculated according to the following formula d :

[0096] T d =k t pAD (10)

[0097] where k t Represents the wind moment coefficient, p is the wind pressure, p=1 / 2ρv 2, ρ is the air density, A is the cross-sectional area of ​​the turntable servo system, which is regarded as a prototype, then A=πD 2 / 4, D is the diameter of the approximately circular cross-section of the turntable servo system.

[0098] Step 22: Add the wind interference model (Equation (10)) to Equation (3) to obtain the second servo system model with gust interference:

[0099]

[0100] Step 3: Generate training samples.

[0101] Since it is very difficult to directly collect the gust interference of the turntable through sensors, the inventors use an indirect method to calculate and obtain it as a training set.

[0102] In this step, the first servo system model and the second servo system model generate system state data based on the same control variable input, and the wind torque T is calculated based on the difference between the system state data generated by the two servo system models. d ; The system state data generated by the second servo system model containing gust interference and the calculated wind torque T d , forming training samples.

[0103] Build 2 sets of Figure 2 The control system shown in FIG. 1 uses the first servo system model and the second servo system model respectively. When obtaining the training sample, the PID control algorithm can be used to give the desired position x 1d The sine wave is used as the tracking signal, and the PID control algorithm generates the control quantity u, which is input to the first servo system model and the second servo system model at the same time. The two models each output the angular position q, and the angular velocity is calculated. In order to distinguish the two models, the output angular position of the first servo system model is recorded as q1, and the angular velocity is calculated as The output angular position of the second servo system model is recorded as q2, and the calculated angular velocity is The wind moment can be calculated based on the data generated by the two models

[0104] The training samples are: q1, T d Among them, q1, is the neural network input, T d is the output of the neural network. The T generated in this step d Labels to use for the learning process.

[0105] Step 4: Use the training samples to train the spectral normalization neural network to have wind moment T d estimation ability.

[0106] The input of the spectral normalization neural network is q1, Output is T d During training, the input q1, From the first servo system model with wind disturbance; in actual control, q1, Collected from the actual turntable servo system.

[0107] The spectral normalization algorithm stabilizes the training of neural networks by limiting the Lipschitz constant of the objective function, which helps to improve the generalization performance of neural networks.

[0108] For the nth layer of the neural network, the input x n-1 With the output x n The relationship can be expressed as follows:

[0109] x n =a n (W n x n-1 +b n ) (11)

[0110] where a n (·) is the activation function of the neural network in this layer, where ReLU function is used, W n is the network parameter matrix, b n is the bias of the network. For the convenience of derivation, b n If we omit the above, equation (12) can be written as

[0111] x n =D n W n x n-1 (12)

[0112] Among them D n ReLU represents the function of the ReLU function, which is a diagonal matrix. When the input value is less than 0, the diagonal element is 0; when the input value is greater than 0, the diagonal element is 1. The ReLU neural network can be described as

[0113] f(x)=D N W N ···D1W1x (13)

[0114] The Lipschitz constraint places requirements on the gradient of f(x)

[0115]

[0116] where ||W i ||2 represents the spectral norm of the parameter matrix. For the diagonal matrix D n, the spectral norm is the maximum value of the diagonal elements. Since the diagonal elements corresponding to ReLU are 0 or 1, its spectral norm is less than or equal to 1. Therefore, the above formula can be expressed as

[0117]

[0118] In order to make f(x) satisfy the Lipschitz constraint, the above formula is spectrally normalized to obtain

[0119]

[0120] Simply put, by dividing the parameter matrix of each layer of the neural network by its spectral norm, the Lipschitz constant can be ensured to be less than or equal to 1. If the constant is to be less than or equal to the set hyperparameter γ, each parameter matrix W can be i The following spectral normalization algorithm is used:

[0121]

[0122] where γ is the Lipschitz constant expected by the neural network, N+1 is the number of neural network layers, and σ(·) is the spectral norm.

[0123] Step 5: Construct a preset performance sliding mode controller: The tracking error e of the turntable servo system is first converted into a conversion error ε according to the preset performance function, and then the conversion error ε is used for sliding mode control to output the control variable u.

[0124] The preset performance control goal is to ensure that the tracking error meets the preset transient performance and steady-state performance during the control process of the turntable servo system, that is,

[0125] -ρ(t)<e(t)<ρ(t) (18)where e(t)=y(t)-y d (t) is the tracking error, ρ(t)=(ρ0-ρ ∞ )e -at +ρ ∞ is the performance function. ρ0 is the initial value range of the tracking error specified by the performance function, ρ ∞ The steady-state boundary of the tracking error when the time t specified by the performance function approaches positive infinity; a is the attenuation factor.

[0126] According to the performance boundary ρ(t), the conversion error is designed to be

[0127]

[0128] Where λ(t) = e(t) / ρ(t).

[0129] Derivative of the conversion error ε gives

[0130]

[0131] Select the linear sliding surface

[0132]

[0133] Where c is an adjustable gain greater than zero, and the first-order derivative of the sliding surface is

[0134]

[0135] According to formula (22),

[0136]

[0137] in x 1d is the desired angular position.

[0138] Finally, the preset performance sliding mode controller based on spectral normalized neural network is designed as

[0139]

[0140] Where η and k are both constants greater than zero, sgn(·) is the sign function, The spectral normalization neural network is used to calculate the gust disturbance T d The approximation result of .

[0141] Step 6: During actual control, the wind torque estimation is obtained using the spectral normalization neural network according to the system status data output by the turntable servo system. Estimate the wind moment The compensation is fed into the preset performance sliding mode controller to generate the control variable u.

[0142] The control system architecture remains the same Figure 2 As shown in the figure, the position of the turntable servo system is the actual turntable servo system. The angular position and angular velocity are collected from the actual turntable and input into the trained spectrum normalization neural network to obtain the wind torque estimation. Formula (24) is used to generate the control variable u, which is output to the turntable servo system for continuous control.

[0143] This concludes the process.

[0144] The technical solution disclosed in the present invention is simulated and verified as follows:

[0145] S1: In the mathematical model of the turntable servo system of the simulation object, J=0.1, K1=1, K2=0.2.

[0146] S2: The simulation parameters for wind load modeling are as follows: Steady-state wind speed v m =10m / s, axial wind force coefficient k A =1.2, the side wind moment coefficient is kc =0.4, wind moment coefficient k t =0.16, radius D = 0.5m, sampling time 0.1s, wind torque simulation results are as follows Figure 5 shown.

[0147] S3: The present invention uses the gust interference model to obtain data and adopts a 5th-order autoregressive model to approximate the wind interference. The data set is divided into 70% training, 15% verification, and 15% test. Then a 6-layer neural network (5-10-16-20-8-1 neurons) is constructed, and the system angle and angular velocity are input to predict the wind interference torque. Finally, the random gradient descent method is used to optimize the network parameters, and the verification set is used to prevent overfitting, and the test set is used to evaluate the effect. Finally, the trained neural network model is saved with the training results as shown in the figure. Figure 6 shown.

[0148] S4: The preset performance parameters are ρ(0)=2, ρ(∞)=0.05, α=6, the sliding mode parameter is c=15, and the reaching law parameter is η=5, k=8. Gust interference T d Use spectral normalization to normalize the neural network output Approximation. The expected tracking signal is a sinusoidal signal y d = sin(πt). The initial value is designed to be x(0) = [1,0] T The simulation time is 10s. The simulation results are as follows Figure 7-9 shown.

[0149] Figure 7 and Figure 8 They are respectively the position tracking and position error and the speed tracking and speed error of the system under the preset performance sliding mode control based on the spectral normalized neural network. The results show that the error quickly approaches zero after 0.5 seconds, and the control effect is good. Figure 9 In the control process of the preset performance sliding mode controller, the spectral normalized neural network approximates the wind disturbance. It can be seen that the approximation value and the true value curve are approximately coincident, indicating that the neural network has a good approximation effect.

[0150] In order to verify the superiority of the proposed compensation algorithm based on spectral normalization neural network, this paper also carried out a preset performance sliding mode control simulation experiment without using spectral normalization neural network compensation as a comparison. Figure 10 and Figure 11 They are the position tracking and position error and the velocity tracking and velocity error of the system under the preset performance sliding mode control without compensation.

[0151] Figure 10 、 Figure 11 and Figure 7 、 Figure 8Comparison shows that the position error of the system under the preset performance sliding mode control with compensation and the system under the preset performance sliding mode control without compensation are both within the performance function constraints, but the system position error is reduced compared to the uncompensated case, thus verifying the effectiveness of the proposed compensation algorithm. In addition, to verify the superiority of the preset performance sliding mode control algorithm, we conducted a simulation experiment comparing it with traditional sliding mode control. Figure 12 and Figure 13 They are the position tracking and position error and the velocity tracking and velocity error of the system under sliding mode control. Figure 10 、 Figure 11 By comparison, it can be seen that under the preset performance sliding mode control, the system's position tracking performance and speed tracking performance are better, and both the control accuracy and dynamic response speed are improved.

[0152] The above specific embodiments merely illustrate the design principles of the present invention. The shapes and names of the components described herein may vary and are not limiting. Therefore, those skilled in the art may modify or substitute equivalents for the technical solutions described in the above embodiments. Such modifications and substitutions, without departing from the inventive spirit and technical solutions of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A servo system preset performance sliding mode control method based on spectral normalization neural network, characterized in that: include: Step 1: For the turntable servo system, build a disturbance-free first servo system model; Step 2: Construct a wind disturbance model and convert the wind moment output by the wind disturbance model into Add the first servo system model to obtain the second servo system model including gust interference; In the wind interference model, the wind speed Steady-state wind speed and fluctuating wind speed sum; The steady-state wind speed ;in, is the steady-state wind speed at height ℎ, Reference height The steady-state wind speed at is the surface roughness index; The pulsating wind speed The way to obtain it is: construct a filter function H , so that its frequency response curve is close to the Davenport spectrum curve; white noise with a mean of zero is passed through this filter function H , filter function H Output pulsating wind speed ; The wind speed The model is substituted into the following wind moment Calculation formula to obtain wind interference model: in, represents the wind moment coefficient, is the wind pressure, , is the air density, A is the area of ​​the circular cross section of the turntable servo system, D is the diameter of the circular cross section of the turntable servo system; Step 3: The first servo system model and the second servo system model each generate system state data based on the same control variable input, and the wind torque is calculated based on the difference between the system state data generated by the two servo system models. ; The system state data generated by the second servo system model containing gust interference and the calculated wind torque , forming training samples; Step 4: Use the training samples to train the spectral normalization neural network to have wind moment Estimation ability; Step 5: Build a Sliding Mode Controller with Preset Performance: Tracking Error of the Turntable Servo System First generate the conversion error according to the preset performance function , and then use the conversion error Perform sliding mode control and output control quantity u ; Among them, the conversion error for: in, , is the tracking error of the turntable servo system, for is the performance function, The initial range of tracking error specified for the performance function, Time specified for performance functions t The steady-state bound of the tracking error as it approaches positive infinity; is the attenuation factor; the resulting conversion error as the error in the sliding surface; The sliding surface is a linear sliding surface. : ;in, is an adjustable gain greater than zero; The preset performance sliding mode controller constructed is: In the above formula, , is the angular position, is the angular velocity, For a given angle, J is the motor inertia, is the system load torque; and are the parameters of the sliding surface, all of which are constants greater than zero; is a symbolic function; , , is the torque time constant, is the rotor resistance, , is the number of electrodes, is the rotor flux; Step 6: During actual control, the wind torque estimation is obtained using the spectral normalization neural network according to the system status data output by the turntable servo system. ; Estimate the wind moment Compensate to the preset performance sliding mode controller to generate the control quantity u .

2. The method according to claim 1, wherein The disturbance-free first servo system model is: The second servo system model with gust interference is: in, , q is the angular position, is the angular velocity; J is the motor inertia; , , ; is the rotor resistance, is the torque time constant, is the system load torque, is the number of electrodes, is the rotor flux; y is the angular position of the turntable servo system.

3. The method according to claim 1, wherein The filter function H for: in, is the Laplace operator.

4. The method according to claim 1, wherein The input of the spectral normalization neural network is the system state data of the turntable servo system, including the angular position q and angular velocity .

5. The method according to claim 1, wherein In step 3, the wind torque is calculated based on the difference of the system state data generated by the two servo system models. for; Get the output angular velocity of the two servo system models The difference , calculate the wind moment .

Images