Satellite attitude super-agile sweep neural network minimum parameter learning sliding mode control method

Abstract

The invention discloses a satellite attitude super-agile sweep neural network minimum parameter learning sliding mode control method, which belongs to the technical field of microsatellite attitude control, and comprises the following steps: S1, in a satellite control process, establishing an actual output model of a neural network; s2, introducing a minimum parameter learning method to replace weight adjustment of a radial basis function neural network, and converting calculation of a neural network weight matrix W into solution of a single parameter # imgabs0 #; according to the method, the system state is converged to zero within finite time by using fast Terminal sliding mode control, and a switching item in a sliding mode is approached through a radial basis function neural network, so that the aim of weakening buffeting is fulfilled; meanwhile, a minimum parameter learning method is designed to replace weight adjustment of the radial basis function neural network, a large amount of operation of a weight matrix is avoided, the efficiency of the controller is effectively improved, and the method has the advantages of being good in convergence, imaging and control effect.

Description

Technical Field

[0001] The present invention relates to the technical field of micro-satellite attitude control, and specifically to a satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method. Background Art

[0002] High-performance and low-cost micro-satellite platforms are one of the future directions of aerospace development. As the core part of micro-satellites, attitude control systems are the most important part of the development of general commercial satellite platforms. At the same time, attitude control systems are also the part of satellites that have the most failures. Therefore, it is very important to maximize the reliability of attitude and orbit control subsystems within the constraints of small satellite size, mass, power, and cost.

[0003] The magnetic torquer is a commonly used attitude control actuator. It uses the electric energy on the satellite to generate a dipole magnetic moment through the coil on the magnetic bar. The geomagnetic field interacts with the generated magnetic moment to generate a magnetic control torque to control the satellite's attitude. It has the advantages of not requiring the consumption of working fluid, low power consumption, low cost, and small volume and mass. It has been used on about two-thirds of satellites. Satellites are subject to various torques in orbit, which will affect the satellite's attitude and further affect the normal operation of the satellite. Since the magnetic torquer needs to rely on the geomagnetic field to generate control torque, only using the magnetic torquer for active control cannot fully control the three axes of the satellite in real time. Therefore, the control results of different control algorithms vary greatly. How to design a good control algorithm is the main problem facing satellite active magnetic control.

[0004] Therefore, in order to ensure that the high-quality imaging of the ultra-agile sweeping maneuvers of low-orbit microsatellites can be seamlessly spliced ​​to achieve full detection of regional targets, the satellite attitude controller is required to achieve high-precision tracking of the planned attitude trajectory and rapid error convergence, and at the same time, it is required to have the ability to stably point to the target. Currently, there are not many studies on ultra-agile maneuvering control of microsatellites, and in the research on attitude control algorithms for other application scenarios, generally only one or two of the tracking accuracy, convergence speed and stable pointing are focused on, which makes it difficult to meet the control requirements of ultra-agile maneuvering and stable imaging. For the application scenarios of this patent, in order to achieve ultra-agile attitude maneuvers of microsatellites and ensure the quality of tracking imaging, the attitude control algorithm is required to have high-precision tracking, rapid convergence and stable pointing capabilities.

[0005] Therefore, in view of the above problems, it is urgent to design a satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method to solve the problems existing in the above-mentioned prior art. Summary of the invention

[0006] In view of the above-mentioned problems, the present invention aims to provide a satellite attitude ultra-agile swing neural network minimum parameter learning sliding mode control method. This method makes the system state converge to zero within a finite time by utilizing fast Terminal sliding mode control, and approximates the switching terms in the sliding mode by using a radial basis function neural network, thereby achieving the purpose of weakening the jitter. At the same time, a minimum parameter learning method is designed to replace the weight adjustment of the radial basis function neural network, avoiding a large number of weight matrix calculations, effectively improving the efficiency of the controller, and having the characteristics of good convergence, imaging and control effects.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method, comprising the steps of

[0009] S1. Establish the actual output model of the neural network during satellite control;

[0010] S2. Based on the actual output model of the neural network, the minimum parameter learning method is introduced to replace the weight adjustment of the radial basis function neural network, and the calculation of the neural network weight matrix W is converted into the calculation of a single parameter By solving the problem, the symbolic function term sgn(s) in the sliding mode controller can be approximated.

[0011] Furthermore, the process of establishing the actual output model of the radial basis function neural network described in step S1 includes:

[0012] (1) In the satellite control process, the estimated value of the sign function of the sliding surface s output by the sliding mode controller is:

[0013] u sgn =ρf(x) (1)

[0014] Where f(x) is the ideal output of the symbolic term approximated by the radial basis function neural network;

[0015] (2) Yes

[0016] f=W *T h(x)+e (2)

[0017] Where x = [q ev ω e ] T is the input of the neural network, h represents the Gaussian kernel function, W * represents the ideal value of the weight from the hidden layer to the output layer, and e represents the error of the neural network approximation;

[0018] (3) The actual output of the neural network is:

[0019]

[0020] in, are the estimated weights of the neural network.

[0021] Furthermore, the process of introducing the minimum parameter learning method to replace the weight adjustment of the radial basis function neural network based on the actual output model of the neural network in step S2 includes:

[0022] (1) According to the neural network minimum parameter learning method, let φ is a constant, and φ>0, let Represents the estimate of φ, and the error is:

[0023]

[0024] (2) The new control law after introducing the radial basis function neural network based on the minimum parameter learning method is:

[0025]

[0026] The robust term Ksgn(s) in the control law is used to overcome disturbances and neural network approximation errors to ensure system stability.

[0027]

[0028] Where W is the weight matrix of the neural network, h is the Gaussian kernel function, e is the approximation error of the neural network, and I 3×3 is a 3×3 unit matrix, α, β, g, h, m, n are all sliding surface design parameters, q ev is the error quaternion, ω e is the error angular velocity, J is the satellite moment of inertia, ω b is the satellite body rotation angular velocity, ω d is the desired angular velocity, H CMG To control the angular momentum of the CMG, is the transformation matrix from the desired coordinate system to the body coordinate system, u is the control input, and d0 is the disturbance torque.

[0029] Furthermore, the sliding mode control method also includes step S3. a process of proving the stability of a new control law based on a radial basis function neural network using a minimum parameter learning method.

[0030] Furthermore, the process of proving the stability of the new control law after the radial basis function neural network based on the minimum parameter learning method described in step S3 includes:

[0031] S31. Define the Lyapunov function

[0032]

[0033] Among them, γ>0;

[0034] Taking the derivative of V, we get:

[0035]

[0036] S32. Substitute the control law of formula (1) into formula (5) to obtain:

[0037]

[0038] roll out:

[0039]

[0040] Then we get:

[0041]

[0042] in, Derived The adaptive law is taken as:

[0043]

[0044] Where, κ>0;

[0045] have:

[0046]

[0047] Pick Substituting into formula (8), we have

[0048]

[0049] In the formula,

[0050] Solve this inequality:

[0051]

[0052]

[0053] Therefore, the convergence accuracy of s is jointly determined by φ, γ and η.

[0054] The beneficial effects of the present invention are as follows: the present invention discloses a satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method. Compared with the prior art, the present invention has the following improvements:

[0055] 1. The present invention designs a satellite attitude ultra-agile swing sweep neural network minimum parameter learning sliding mode control method. This method uses minimum parameter learning to adaptively estimate parameters to replace the radial basis function neural network weight adjustment, and converts the calculation of the neural network weight matrix W into the calculation of a single parameter This adaptive solution is simpler, has a shorter calculation time, and is more efficient. It can also achieve the best approximation effect of the radial basis function neural network, and is also conducive to its application in actual satellite control systems.

[0056] 2. This method uses fast Terminal sliding mode control to make the system state converge to zero in a finite time, and uses radial basis function neural network to approximate the switching terms in the sliding mode, thereby achieving the purpose of weakening chattering;

[0057] 3. This method replaces the weight adjustment of the radial basis function neural network by designing a minimum parameter learning method, avoiding a large number of weight matrix calculations, effectively improving the efficiency of the controller, and has the advantages of good convergence, imaging and control effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a block diagram of the neural sliding mode control based on minimum parameter learning of the present invention.

[0059] Figure 2 This is a comparison chart of the desired attitude quaternion of the present invention and the attitude quaternion tracked by the radial basis function neural sliding mode controller.

[0060] Figure 3 This is a graph showing the posture error tracking results of the radial basis function neural sliding mode controller of the present invention.

[0061] Figure 4 This is a diagram showing the angular velocity tracking results of the radial basis function neural sliding mode control of the present invention. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0063] Example 1: Refer to the attached Figure 1-4 A satellite attitude ultra-agile sweep neural network minimum parameter learning sliding mode control method is shown. This method uses the characteristic of radial basis function neural network that can infinitely approximate any nonlinear function, and uses radial basis function neural network to realize the approximation of the symbolic function term sgn(s) in the sliding mode controller to ensure that the switching gain of the sliding mode surface is reduced and the chattering of the sliding mode controller is weakened; the method includes the steps of

[0064] S1. Establishing the actual output model of radial basis function neural network in the satellite control process

[0065] Assume that the estimated value of the sign function of the sliding surface s output by the sliding mode controller is:

[0066] u sgn =ρf(x) (1)

[0067] Where f(x) is the ideal output of the radial basis function neural network approximating the symbolic term, then:

[0068] f=W *T h(x)+e (2)

[0069] Where x = [q ev ω e ] T is the input of the neural network, h represents the Gaussian kernel function, W * represents the ideal value of the weight from the hidden layer to the output layer, and e represents the error of the neural network approximation;

[0070] The actual output of the neural network is:

[0071]

[0072] in, is the estimated weight of the neural network;

[0073] S2. Based on the actual output model of the neural network, the minimum parameter learning method is introduced to replace the weight adjustment of the radial basis function neural network, and the calculation of the neural network weight matrix W is converted into the calculation of a single parameter Solution

[0074] According to the neural network minimum parameter learning method, let φ is a constant, and φ>0, let represents the estimate of φ, then the error can be expressed as:

[0075]

[0076] The new control law after introducing the radial basis function neural network based on the minimum parameter learning method can be expressed as:

[0077]

[0078] The robust term Ksgn(s) in the control law is used to overcome disturbances and neural network approximation errors to ensure system stability.

[0079]

[0080] S3. Prove the stability of the new control law based on the radial basis function neural network based on the minimum parameter learning method shown in equation (6)

[0081] S31. Define the Lyapunov function

[0082]

[0083] Among them, γ>0;

[0084] Taking the derivative of V, we can get:

[0085]

[0086] S32. Substituting the control law of formula (1) into formula (5), we can get:

[0087]

[0088] It can be introduced:

[0089]

[0090] Then we can get:

[0091]

[0092] in, And it can be deduced that

[0093] The adaptive law is taken as:

[0094]

[0095] In the formula, κ>0.

[0096] So we have:

[0097]

[0098] Pick Substituting into formula (8), we have

[0099]

[0100] In the formula,

[0101] Solve this inequality:

[0102]

[0103] It can be seen that the convergence accuracy of s is jointly determined by φ, γ and η.

[0104] Embodiment 2: Different from the embodiment, in order to verify the effectiveness of the control algorithm proposed in embodiment 1, this embodiment is designed to simulate and verify the ultra-agile swing-sweep sliding mode control method of micro-satellite attitude based on minimum parameter learning adaptive radial basis function neural network proposed in this patent. The basic conditions of the simulation are set as follows, Table 1 is the satellite parameter table, and Table 2 is the controller parameter table.

[0105] Table 1: Satellite and orbit parameters in radial basis function sliding mode control

[0106]

[0107] Table 2: Radial basis function sliding mode controller parameters

[0108]

[0109] The simulation results are as follows Figure 2-4 As shown;

[0110] The proposed ultra-agile sweep sliding mode control method for micro-satellite attitude based on minimum parameter learning adaptive radial basis function neural network is simulated and verified. Figure 2 The middle figure (a) and figure (b) respectively represent the planned expected posture quaternion and the posture quaternion result obtained by tracking the expected posture under the control algorithm proposed in this patent. The expected posture and the tracking posture curve still maintain the same change trend.

[0111] By calculating the error between the desired attitude quaternion and the attitude tracking quaternion, we can get Figure 3 Figure (a) shows the attitude tracking error expressed by Euler angles, and Figure (b) shows the attitude tracking error expressed by quaternions. The maximum attitude angle errors on the three axes are 0.009°, 0.008°, and 0.009°, respectively. The error accuracy is basically within 10 -3 About order of magnitude. Figure 4 The tracking error results of the angular velocity of the ultra-agile sweep sliding mode control method for micro-satellite attitude based on the minimum parameter learning adaptive radial basis function neural network are shown. The maximum angular velocity errors on the three axes are 0.14° / s, 0.17° / s and 0.17° / s respectively. The simulation results verify that the algorithm proposed in Example 1 of the present invention can meet the control requirements of ultra-agile maneuverable stable imaging of micro-satellite attitude.

[0112] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method, characterized by: Included steps S1. Establish the actual output model of the neural network during satellite control; S2. Based on the actual output model of the neural network, the minimum parameter learning method is introduced to replace the weight adjustment of the radial basis function neural network, and the calculation of the neural network weight matrix W is converted into the calculation of a single parameter By solving the problem, the symbolic function term sgn(s) in the sliding mode controller can be approximated.

2. A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method as claimed in claim 1, characterized in that: The process of establishing the actual output model of the radial basis function neural network described in step S1 includes: (1) In the satellite control process, the estimated value of the sign function of the sliding surface s output by the sliding mode controller is: u sgn =ρf(x) (1) Where f(x) is the ideal output of the symbolic term approximated by the radial basis function neural network; (2) Yes f=W *T h(x)+e (2) Where x = [q ev ω e ] T is the input of the neural network, h represents the Gaussian kernel function, W * represents the ideal value of the weight from the hidden layer to the output layer, and e represents the error of the neural network approximation; (3) The actual output of the neural network is: in, are the estimated weights of the neural network.

3. A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method as claimed in claim 2, characterized in that: The process of introducing the minimum parameter learning method to replace the weight adjustment of the radial basis function neural network based on the actual output model of the neural network described in step S2 includes: (1) According to the neural network minimum parameter learning method, let φ is a constant, and φ>0, let Represents the estimate of φ, and the error is: (2) The new control law after introducing the radial basis function neural network based on the minimum parameter learning method is: The robust term Ksgn(s) in the control law is used to overcome disturbances and neural network approximation errors to ensure system stability. Where W is the weight matrix of the neural network, h is the Gaussian kernel function, e is the approximation error of the neural network, and I 3×3 is a 3×3 unit matrix, α, β, g, h, m, n are all sliding surface design parameters, q ev is the error quaternion, ω e is the error angular velocity, J is the satellite moment of inertia, ω b is the satellite body rotation angular velocity, ω d is the desired angular velocity, H CMG To control the angular momentum of the CMG, is the transformation matrix from the desired coordinate system to the body coordinate system, u is the control input, and d0 is the disturbance torque.

4. A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method as claimed in claim 3, characterized in that: The sliding mode control method also includes step S3. a process of proving the stability of a new control law based on a radial basis function neural network using a minimum parameter learning method.

5. A satellite attitude ultra-agile swing-sweep neural network minimum parameter learning sliding mode control method as claimed in claim 4, characterized in that: The process of proving the stability of the new control law after the radial basis function neural network based on the minimum parameter learning method described in step S3 includes: S31. Define the Lyapunov function Among them, γ>0; Taking the derivative of V, we get: S32. Substitute the control law of formula (1) into formula (5) to obtain: roll out: Then we get: in, Derived The adaptive law is taken as: Where, κ>0; have: Pick Substituting into formula (8), we have In the formula, Solve this inequality: Therefore, the convergence accuracy of s is jointly determined by φ, γ and η.

Images