A neural network global fast terminal sliding mode control method for momentum wheels
Through the combination of neural network global fast terminal sliding mode control and RBF neural network observer, the problem of uncertainty and interference in momentum wheel operation is solved, the speed control accuracy and anti-interference ability are improved, and the efficient and stable momentum wheel system is achieved. control.
Patent Information
- Application Number
- CN202211729777.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-30
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-12-30
AI Technical Summary
There are uncertainty and unknown interferences during the momentum wheel during operation, resulting in a decrease in dynamic and static performance of the output speed of the momentum wheel drive motor, making it difficult to improve the control accuracy. Traditional PID control and other intelligent control algorithms have shortcomings in anti-interference.
The sliding mode control method of the neural network is adopted, combined with the RBF neural network observer to estimate and compensate for unknown total interference, and the index function is adjusted through the sliding mode control design weight, and the gradient descent method is used to optimize the neural network parameters to improve the Speed control accuracy and anti-interference ability of the momentum wheel.
The momentum wheel speed tracking error converges to zero in a limited time, improves the speed tracking performance and anti-interference ability, and significantly improves the control accuracy and stability of the momentum wheel system.
Smart Images

Figure CN115774398B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of satellite technology, and in particular relates to a neural network global fast terminal sliding mode control method for a momentum wheel. Background Art
[0002] When a micro-nano satellite is in orbit, its attitude must be controlled to ensure the completion of specific tasks, such as aiming the onboard camera at the observation area and the satellite antenna at the service area. The momentum wheel is an actuator widely used in satellite attitude control systems. According to the momentum theorem, it exchanges momentum with the satellite body to achieve satellite attitude stability and control, so it is also called a momentum exchange actuator.
[0003] At present, micro-nano satellites have moved from simple technology demonstration and verification to functional applications in various fields, which also puts forward higher precision requirements for satellite attitude control. The momentum wheel is essentially a wheel body with large angular momentum that rotates at high speed driven by a motor. Improving the control accuracy of the momentum wheel actually means improving the control accuracy of the momentum wheel drive motor.
[0004] Momentum wheel drive motors at home and abroad all use brushless DC motors. Since the magnetic materials in the motor have hysteresis and saturation phenomena, and the magnetic properties of the permanent magnets change nonlinearly with temperature, the motor has time-varying, nonlinear, and strong coupling characteristics. These characteristics determine that its mathematical model is difficult to establish accurately, and its dynamic parameters cannot be accurately obtained. In addition, there are system uncertainties and unknown interferences during the operation of the momentum wheel, which seriously affect the dynamic and static performance of the motor output speed and reduce the control accuracy of the momentum wheel.
[0005] The traditional PID control method cannot take into account the requirements of the momentum wheel in terms of speed control and anti-interference. The new intelligent control algorithms, such as fuzzy control and genetic algorithm, have the disadvantages of complex structure and difficulty in implementation. Further improvement is needed in the field of momentum wheel motor control. Summary of the invention
[0006] In order to solve the above problems, the purpose of the present invention is to provide a neural network global fast terminal sliding mode control method for a momentum wheel, which is used to solve the problem of how to eliminate uncertainty and unknown interference in the operation of the momentum wheel and improve the speed control accuracy and anti-interference ability of the momentum wheel.
[0007] In order to achieve the above-summarized purpose, the technical solution adopted by the present invention is as follows:
[0008] A neural network global fast terminal sliding mode control method for momentum wheels.
[0009] The final total control rate is:
[0010]
[0011] The current control rate is:
[0012]
[0013] Among them, θ n is the nominal value of the control coefficient of the momentum wheel system, α, β are gain coefficients, p / q is the exponential coefficient, s is the sliding surface of the system, is the derivative of the speed command signal, x2 is the angular acceleration of the motor, is the second-order derivative of the speed command signal, is the estimated value of the total disturbance of the system, sign(s) is the switching function of the sliding surface s, ξ is the function gain term, and γ are gain coefficients, and t is the time variable.
[0014] Compared with the prior art, the present invention has the following advantages:
[0015] (1) The present invention adopts global fast terminal sliding mode control, so that the momentum wheel speed tracking error can converge to zero within a limited time, thereby improving the speed tracking performance of the momentum wheel.
[0016] (2) The present invention uses an RBF neural network observer to estimate the unknown total interference, and adds the estimated value of the unknown total interference to the control quantity for compensation, thereby improving the control performance and enhancing the anti-interference ability of the momentum wheel system.
[0017] (3) The present invention designs a weight adjustment index function in combination with sliding mode control, and uses the gradient descent method on the basis of this function to optimize the parameters of the neural network, thereby further improving the observation ability of the neural network.
[0018] The above-described objects, features and advantages as well as the working principle of the present invention will be described and explained in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a principle block diagram of the momentum wheel motor double closed-loop control system in the present invention;
[0020] Figure 2 It is a schematic diagram of the RBF neural network structure used in the present invention;
[0021] Figure 3 It is a neural network global fast terminal sliding mode control flow chart in the present invention;
[0022] Figure 4 It is the speed tracking comparison chart of ANNFTSMC, SMC and PID control when the acceleration command is given;
[0023] Figure 5 It is the speed tracking comparison chart of ANNFTSMC, SMC and PID control when the deceleration command is given;
[0024] Figure 6 This is the speed comparison chart of ANNFTSMC, SMC and PID control when the system is suddenly disturbed;
[0025] Figure 7 It is the speed curve of ANNFTSMC control after the system parameters are changed;
[0026] Figure 8 It is the total disturbance estimation diagram of ANNFTSMC control after the total disturbance of the system is given. DETAILED DESCRIPTION
[0027] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0028] The present invention is a neural network global fast terminal sliding mode control method for a momentum wheel. The above technical solution is specifically implemented by the following steps:
[0029] Step 1: Establish a dual closed-loop control topology of the momentum wheel brushless DC motor.
[0030] like Figure 1 As shown in the figure, the dual closed-loop control topology of the momentum wheel brushless DC motor includes: speed controller, current controller, three-phase full-bridge inverter, momentum wheel, Hall position sensor, commutation calculation module, etc. The speed loop is designed with a neural network global fast terminal sliding mode controller (ANNFTSMC), and the current loop adopts a proportional integral controller (PI).
[0031] Step 2: Establish a mathematical model of the momentum wheel based on the topological structure.
[0032] According to the above topological structure, the momentum wheel DC brushless motor system including the current loop is mathematically modeled, and its electromagnetic torque equation is:
[0033]
[0034] Where: T e is the electromagnetic torque of the motor, i a ,i b ,i c is the three-phase current of the stator winding, e a ,e b ,e cis the three-phase back electromotive force of the stator winding, and ω is the rotor angular velocity. When the momentum wheel operates in the 120° conduction mode and the commutation process is not considered, only two phases of the three-phase star-shaped stator winding are turned on, and their currents are equal in magnitude and opposite in direction. In this way, the above formula can be simplified to:
[0035] T e =K t i (2)
[0036] Where: K t is the motor torque coefficient, i is the winding phase current in steady state. Ignoring the influence of viscous friction coefficient, the motor motion equation is:
[0037]
[0038] Where: T x (t) is the unknown disturbance torque of the momentum wheel system, which changes with time, J is the moment of inertia of the momentum wheel, which is the sum of the moment of inertia of the motor and the moment of inertia of the mass block, ω is the angular velocity of the motor, and t is the running time of the motor.
[0039] To facilitate the design of the controller, the state variable x1 = ω is selected. Then the state equation of the momentum wheel motor system can be expressed as:
[0040]
[0041] Where: x1, x2 are the angular velocity and angular acceleration of the motor, which are the state variables of the system. is the control quantity of the system, y is the output of the system, θ=K t / J is the coefficient of the system control quantity, which is related to the system parameters. is the unknown disturbance of the system. To facilitate the subsequent controller design, the following assumptions are made:
[0042] Assumption 1: The unknown disturbance d(t) to the controller is bounded, and there exists a positive number ξ that satisfies |d(t)|≤ξ.
[0043] Assumption 2: The coefficient θ of the control variable is bounded during actual operation.
[0044] Step 3: Considering the uncertainty and unknown disturbance in the operation of the momentum wheel, the global fast terminal sliding mode control method is used to design the controller.
[0045] When the momentum wheel is working normally, a speed command signal y is given. d When t→∞, the system will output signal y to track the speed command signal. The control target is: when t→∞, (yy d )→0, where the system output signal y=x1, and the speed command signal y d=x d .
[0046] Assumption 3: The system speed command signal x d is bounded.
[0047] Let the speed tracking error be:
[0048] e=x d -x1 (5)
[0049] Design sliding surface s:
[0050] s=e+αe+βe p / q (6)
[0051] Where: α, β are gain coefficients and α>0, β>0, p / q is the exponential coefficient, p and q are odd numbers and 0<p<q, and the derivative of s is:
[0052]
[0053] make At the same time, substituting formula (4) into formula (7), we get:
[0054]
[0055] Taking into account the uncertainty and unknown disturbance of the control system in the actual operation process, the control quantity u of the system can be divided into an equivalent term u1 and a robust term u2. According to the arrival condition of the sliding surface s = 0, let The equivalent term u1 is:
[0056]
[0057] The robust term u2 of the designed control quantity is:
[0058]
[0059] Then the total control quantity u is:
[0060] u=u1+u2 (11)
[0061] Where: sign(s) is the switching function of the sliding surface s, and γ are gain coefficients and ξ is the function gain term and ξ>0, its value is the same as in Assumption 1. Substituting equation (11) into equation (8), we can get:
[0062]
[0063] Lemma 1: Assume that there exists a continuous positive definite function V(t)>0. If we want to sTo make V(t) converge to 0, the following inequality must be satisfied:
[0064]
[0065] Where: μ, λ, ζ are coefficients and μ>0, λ>0, 0<ζ<1, t0 is the initial time, and:
[0066]
[0067] Stability analysis of global fast terminal sliding mode controller:
[0068] The Lyapunov function is designed as follows:
[0069]
[0070] Deriving equation (15) and substituting equation (12) into it, we get:
[0071]
[0072] Simplifying the above formula, we can get:
[0073]
[0074] According to Lemma 1, the system can be on the designed sliding surface s for a finite time t s The momentum wheel control system can achieve finite time stability with inner convergence.
[0075] Step 4: Use a neural network observer to observe and compensate for the uncertainty and unknown disturbance of the momentum wheel system.
[0076] Step 4.1: Establish the state equation of the momentum wheel including system uncertainty and unknown disturbance
[0077] Since the system parameters cannot be accurately measured, the nominal value of the system control coefficient is defined as θ n , the coefficient uncertainty is Δθ, and we can get θ=θ n +Δθ. The momentum wheel state equation can be expressed as:
[0078]
[0079] In order to facilitate the processing of uncertainties and unknown disturbances in the system, the total disturbance of the system can be:
[0080] D=Δθu+d(t) (19)
[0081] Substituting formula (19) into formula (18), we can obtain:
[0082]
[0083] Step 4.2: Design the control variable containing the neural network observer estimate
[0084] In the embodiment of the present invention, the sliding surface still uses the sliding surface s designed in the global fast terminal sliding mode controller, and the RBF neural network observer is used to observe the total disturbance D. According to the approximation characteristics of the neural network, there is a weight vector W T , so that:
[0085] D=W T h i (x)+ε (21)
[0086]
[0087] Where: i = 1, ..., m, h i (x)=[h1(x),h2(x),h3(x),…,h m (x)] T is the Gaussian radial basis function output of the network, m is the number of hidden layer nodes of the network, x=[x1,…,x n ] T is the network input, n is the number of inputs, W T is the ideal weight, ε is the approximation error, c i is the center value of the radial basis function, c i =[c i1 ,c i2 ,…,c in ],σ i is the width of the radial basis function, σ i =[σ1,σ2,…,σ m ] The structure of the RBF neural network in the present invention is as follows Figure 2 As shown in the figure, the number of hidden layer nodes is set to 5.
[0088] Since the weight vector W of the neural network is unknown, let As the estimated value of W, the estimated value of the total disturbance can be The control quantities of the redesigned control system are as follows:
[0089]
[0090]
[0091]
[0092] Where: γ is the gain coefficient and p / q is the exponential coefficient, p and q are odd numbers and 0<p<q, u3 and u4 are the equivalent term and robust term of the new control quantity respectively.
[0093] Step 4.3: Adjust the weights of the neural network and the center value and width of the basis function to obtain the final total control rate and current control rate
[0094] In order to enable the neural network to accurately observe the total disturbance, it is necessary to adjust the network weights and basis function parameters to achieve the optimal value. From formula (16), it can be seen that when the sliding mode control goal is achieved Therefore, the weight adjustment index function of the neural network can be taken as:
[0095]
[0096] According to the gradient descent method, the update increments of the network weights and the center value and width of the Gaussian basis function are as follows:
[0097]
[0098]
[0099]
[0100] Where: i is the number of network input, j is the number of network hidden layer, is the sign of partial derivative. From equation (8), equation (22) and equation (24), we get:
[0101]
[0102]
[0103]
[0104]
[0105]
[0106] From the above formula, we can see that the update rate of network weights and the center value and width of the Gaussian basis function is as follows:
[0107]
[0108] c ij (t) = c ij (t-1) +Δc ij +a[c ij (t-1)-c ij (t-2)] (36)
[0109] σ j (t) = σ j (t-1)+Δσ j +a(σ j (t-1)-σj (t-2)) (37)
[0110] In the formula, η is the learning rate and a is the momentum factor, and their value range is (0,1).
[0111] In summary, the final total control rate of the present invention is:
[0112]
[0113] The current control rate is:
[0114]
[0115] The neural network global fast terminal sliding mode control method for momentum wheel is simulated and verified.
[0116] The parameters of the momentum wheel DC brushless motor are: rated voltage U = 12V, rated current I = 2.02A, momentum wheel moment of inertia J = 1.4e -4 kg / m 2 , motor viscous friction coefficient B = 0N.ms, phase resistance R = 1.2Ω, phase inductance H = 0.56mH, torque constant K t =25.5mNm / A.
[0117] The current loop PI controller parameter is K p =0.5,K i =0.01. The parameters of the neural network global terminal sliding mode controller are set as:
[0118] α=50,β=2,p=7,q=9, γ=1,ξ=0.5
[0119] The learning rate η in the gradient descent method is 0.2, and the momentum factor a is 0.15.
[0120] Figure 4 This is a speed acceleration tracking comparison of neural network global fast terminal sliding mode control (ANNFTSMC), traditional sliding mode control (SMC) and PID control. It can be seen from the figure that the speed overshoot under traditional PID control is about 78rpm, and there is basically no overshoot in the speed tracking under the two sliding mode controls. In addition, the speed stabilization time under ANNFTSMC control is shorter, about 0.1s, which indicates that this control method has good speed tracking performance.
[0121] Figure 5This is a comparison of the speed deceleration tracking of the neural network global fast terminal sliding mode control, traditional sliding mode control and PID control. It can be seen from the figure that the speed overshoot under the traditional PID control is about 40rpm, and there is basically no overshoot in the speed tracking under the two sliding mode controls, and the speed stabilization time under ANNFTSMC control is shorter, which indicates that the speed tracking performance of ANNFTSMC is better than that of PID and SMC.
[0122] Figure 6 The figure shows a speed comparison of the three control methods when the given speed is 1000rpm and a sudden interference torque of 0.5Nm is added in 1s. It can be seen from the figure that the speed fluctuates significantly under PID and SMC control. The PID control overshoots by about 50rpm and the SMC overshoots by about 10rpm, and it is difficult to recover in a short time. However, the speed does not fluctuate basically under ANNFTSMC control, which shows that this method has good anti-interference ability.
[0123] Figure 7 After setting the moment of inertia of the momentum wheel to twice its original value, the speed curve under ANNFTSMC control is adopted. It can be seen from the figure that the speed tracking effect is still very good, there is basically no overshoot and the stabilization time is very fast. The speed fluctuation is almost 0rpm in steady state, which shows that this control method is not sensitive to changes in the parameters of the momentum wheel system and is suitable for long-term application.
[0124] Figure 8 This is the total disturbance estimation diagram of ANNFTSMC under a given speed of 1000rpm and an additional disturbance torque of 0.05cos(10t)Nm in 1s. It can be seen from the figure that the RBF neural network observer has a relatively accurate estimate of the total disturbance of the system, can compensate for the system control quantity, and can effectively improve the anti-interference ability of the momentum wheel system.
[0125] In summary, the present invention designs a global fast terminal sliding mode controller in the speed control of the momentum wheel, which effectively improves the speed tracking accuracy of the momentum wheel. At the same time, the RBF neural network observer is used to observe the uncertainty and unknown interference of the system, and the observation value is added to the system control quantity for compensation, which effectively improves the anti-interference ability of the momentum wheel control system. Finally, the weight adjustment index function is designed in combination with the sliding mode control. On the basis of this function, the gradient descent method is used to optimize the parameters of the neural network, which further improves the observation ability of the neural network and further improves the anti-interference ability.
Claims
1. A neural network global fast terminal sliding mode control method for a momentum wheel, characterized in that: The following steps are involved: Step 1, establishing a dual closed-loop control topology structure of the momentum wheel brushless DC motor; Step 2: Establish the mathematical model of the momentum wheel based on the topological structure: Where: x1, x2 are the angular velocity and angular acceleration of the motor, is the control quantity of the system, y is the output of the system, θ=K t / J is the coefficient of the system control quantity, is the unknown disturbance of the system, K t is the motor torque coefficient, J is the moment of inertia of the momentum wheel, T x (t) is the unknown disturbance torque of the momentum wheel system; Step 3: Considering the uncertainty and unknown disturbance in the operation of the momentum wheel, the global fast terminal sliding mode control method is used to design the controller; the sliding mode surface s satisfies: Where d(t) is the unknown disturbance of the controller, θ is the coefficient of the system control quantity, u2 is the robust term of the control quantity u; sign(s) is the switching function of the sliding surface s, and γ are adjustment coefficients, and p / q is the exponential coefficient; Step 4: Use a neural network observer to observe and compensate for the uncertainty and unknown disturbance of the momentum wheel system.
2. The neural network global fast terminal sliding mode control method for momentum wheel according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: Establish the state equation of the momentum wheel including system uncertainty and unknown disturbance; Step 4.2: Design the control variable including the estimated value of the neural network observer; Step 4.3: Adjust the weights of the neural network and the center value and width of the basis function to obtain the final total control rate and current control rate.
3. The neural network global fast terminal sliding mode control method for momentum wheel according to claim 2, characterized in that: The momentum wheel state equation established in 4.1 can be expressed as: where θ n is the nominal value of the system control coefficient, and D is the total disturbance of the system.
4. The neural network global fast terminal sliding mode control method for momentum wheel according to claim 3, characterized in that: The system control quantity designed in step 4.2 is: Among them, u3 and u4 are the equivalent term and robust term of the control quantity respectively. γ is the adjustment coefficient.
5. The neural network global fast terminal sliding mode control method for momentum wheel according to claim 1, characterized in that: Step 4.3 gives the final total control rate: The current control rate is: Among them, θ n is the nominal value of the control coefficient of the momentum wheel system, α, β are gain coefficients, p / q is the exponential coefficient, s is the sliding surface of the system, is the derivative of the speed command signal, x2 is the angular acceleration of the motor, is the second-order derivative of the speed command signal, is the estimated value of the total disturbance of the system, sign(s) is the switching function of the sliding surface s, ξ is the function gain term, and γ are gain coefficients, and t is the time variable.
