Cubesat attitude stabilization method based on PID controller parameter setting

Abstract

A CubeSat attitude stabilization method based on PID controller parameter setting combines an inertia weight decline PSO particle swarm optimization algorithm and a variable learning rate BP neural network to carry out PID controller parameter dynamic setting, BP neural network weight parameter initial values suitable for a current CubeSat model are rapidly obtained, PID parameters are dynamically adjusted in a control process by using a BP-PID framework, and the PID parameters are dynamically adjusted in the control process. According to the dynamic parameter setting method under different external disturbance conditions, the requirements for cubesat attitude control performance such as required response time and control precision and the requirement for control robustness are met at the same time.

Description

Technical Field

[0001] This invention relates to a technology in the field of satellite control, specifically a method for stabilizing the attitude of a cubesat based on PID controller parameter tuning. Background Technology

[0002] In the complex and ever-changing space environment, every man-made object in space, such as space stations, spacecraft, and satellites, is susceptible to various disturbances such as solar radiation and external disturbance torques like aerodynamic drag. This can lead to undesirable translational and rotational motions, making attitude control a crucial aspect of spacecraft in-orbit flight. Existing satellite attitude control technologies rely on fixed-parameter PID controllers, which cannot dynamically adjust PID parameters based on real-time attitude errors during control. This results in poor anti-interference capabilities and limited control accuracy when facing complex and varied space disturbances. Furthermore, these technologies lack adaptability to system nonlinearity and external disturbances. Additionally, some technologies employing genetic algorithms suffer from slow convergence speeds, low computational efficiency, and optimization results that get trapped in local optima. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, which are unable to effectively handle long-period attitude swaying and sudden disturbances, lack robustness, and exhibit decreased robustness under sudden disturbances or complex environmental changes. Furthermore, it addresses the difficulty in further optimizing control overshoot and steady-state error. To address these shortcomings, a cubesat attitude stabilization method based on PID controller parameter tuning is proposed. This method combines the inertial weight decay PSO particle swarm optimization algorithm with a variable learning rate BP neural network for dynamic tuning of PID controller parameters. This rapidly obtains initial values ​​of the BP neural network weight parameters suitable for the current cubesat model. The BP-PID framework is then used to dynamically adjust the PID parameters during the control process. This dynamic parameter tuning method, applicable to different external disturbances, simultaneously ensures the required response time, control accuracy, and robustness for cubesat attitude control.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a cubesat attitude stabilization method based on PID controller parameter tuning, comprising:

[0006] Step 1: Based on the known parameters of the CubeSat model and its attitude control system, establish the mathematical expression of its attitude dynamics model based on the attitude expression method of quaternions;

[0007] The aforementioned CubeSat attitude dynamics model employs, but is not limited to, the techniques described by Zhang Renwei in "Satellite Orbit Attitude Dynamics and Control" ([M]. Beijing: Beijing University of Aeronautics and Astronautics Press, 1998: 270-273).

[0008] Step 2: Simulate the closed-loop circuit of the cubesat attitude control system;

[0009] Step 3: Determine the simulation input values ​​and use the PSO optimization algorithm to assign initial weights to each layer of the BP neural network;

[0010] Step 4: Perform forward calculations using the BP neural network and update the PID controller to obtain the control torque;

[0011] Step 5: Update the backpropagation weights of the BP neural network using the gradient descent method;

[0012] Step 6: The required control torque obtained from Step 4 is used to adjust the speed of the momentum wheel through the momentum wheel model to control the attitude of the cube star. After the speed adjustment control is completed, the attitude sensor outputs the current control step cube star attitude.

[0013] The momentum wheel model refers to the actuator of the CubeSat attitude control system, which uses a zero-momentum control scheme to generate a reaction torque (hence also called a reaction wheel) by changing its own rotational speed, which acts on the CubeSat body to achieve attitude control.

[0014] The momentum wheel model described herein is implemented using, but is not limited to, the techniques described by Zhang Renwei in "Satellite Orbital Attitude Dynamics and Control" ([M]. Beijing: Beijing University of Aeronautics and Astronautics Press, 1998: 270-273.).

[0015] Step 7: If the error between the current control step cubesat attitude and the desired attitude does not meet the task requirements, return to step 4 for iteration until the requirements are met.

[0016] This invention relates to a cubesat attitude stabilization system based on PID controller parameter tuning, comprising: an initialization parameter optimization module, an attitude determination module, a control law calculation module, and an actuator drive module. The initialization parameter optimization module performs initial optimization of neural network weights based on inertial weight decay (PSO) and online backpropagation weight update processing based on the system's initial state and attitude error performance indicators during the control process, obtaining BP neural network weight parameter results that optimize system performance. The attitude determination module performs attitude error calculation and signal preprocessing based on noisy Euler angle data measured by attitude sensors and desired attitude command information, obtaining Euler angle error vectors and angular velocity information results for the current attitude and desired attitude. The control law calculation module performs BP neural network forward calculation and PID control law calculation processing based on the attitude error information output by the attitude determination module and the initial weight information provided by the initialization parameter optimization module, obtaining three-axis torque command results for attitude control. The actuator drive module consists of three orthogonally mounted momentum wheels, and performs torque distribution based on the three-axis torque command information output by the control law calculation module to obtain speed adjustment command results for the three zero-momentum wheels.

[0017] Technical effect

[0018] This invention optimizes the initial parameters of a BP neural network based on the PSO (Particle Swarm Optimization) method with inertial weight decay, and then uses this BP neural network to complete real-time parameter tuning during PID control, achieving stable attitude control of a cubesat. Compared with existing technologies, this invention uses a BP network with PSO-optimized initial network parameters for dynamic parameter tuning of the PID controller, resulting in better stability control performance of the cubesat after attitude disturbances, converging to the target attitude with shorter settling time and smaller steady-state error. Furthermore, this invention provides better scenario adaptability and robustness for the cubesat attitude stabilization PID controller when facing complex disturbances. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention;

[0020] Figure 2 This is a schematic diagram of the BP-PID controller structure to be optimized in this invention;

[0021] Figure 3 This is a schematic diagram of the BP neural network structure and forward and backward propagation mechanism used in the embodiment;

[0022] Figure 4 The attitude angle stabilization control response curve of the EM-PID controller used in scenario 1;

[0023] Figure 5The attitude angle stabilization control response curve for scenario 1 using a BP-PID controller;

[0024] Figure 6 The attitude angle stabilization control response curve of the PSO-BP-PID controller used in scenario 1;

[0025] Figure 7 A comparison of the attitude angle stabilization control response curves of the three controllers used in scenario 2. Detailed Implementation

[0026] like Figure 1 As shown, this embodiment relates to a cubesat attitude stabilization method based on PID controller parameter tuning, including:

[0027] Step 1: Based on the known parameters of the CubeSat model and its attitude control system, establish the mathematical expression of its attitude dynamics model based on the attitude expression method of quaternions;

[0028] Step 2: Based on the attitude error expression form of Euler angles, construct a PID controller with a BP neural network, a momentum wheel actuator, and an attitude sensor to simulate the closed loop of the cubesat attitude control system.

[0029] The PID controller outputs the desired control quantity from the current state quantity, and the cubic star momentum wheel, modeled in a three-axis orthogonal manner, independently outputs the torques of the roll, yaw and pitch axes by adjusting each one.

[0030] The BP neural network comprises: an input layer with three neurons, a hidden layer with seven neurons, and an output layer with three neurons. The three neurons in the input layer represent the actual cube star Euler angle value y(k), the reference Euler angle value r(k), and the error e(k) between them, expressed in Euler angle form. The three output neurons correspond to the three parameters K of the PID controller. p K i K d .

[0031] The momentum wheel actuator is a three-axis orthogonally mounted zero-momentum reaction wheel, used to receive three-axis torque commands from the PID controller and independently adjust the speed of each momentum wheel according to the dynamic model, thereby generating the required control torque. Its core function is to provide a standard torque output and execution interface for the control algorithm of this invention.

[0032] The attitude sensor described above is used to simulate the measurement characteristics of a real sensor. Its output consists of real Euler angles calculated by the CubeSat attitude dynamics model superimposed with preset white noise or Gaussian noise. This modeling method aims to verify the robustness of the control algorithm of the present invention in the presence of measurement noise. Its core function is to provide the control loop with an attitude feedback signal containing typical errors.

[0033] Step 3: Based on the attitude perturbation of the CubeSat in actual engineering applications and the reference Euler angle values ​​required by the mission, determine the simulation input values, and use the PSO optimization algorithm to assign the initial weights of each layer of the BP neural network. Specifically, this includes:

[0034] 3.1 Set the total number of particle swarms n; based on the number of network weight parameters D of the BP neural network, treat each set of initialized weight parameters as a particle swarm, and use D-dimensional vectors respectively. and D-dimensional vectors This represents the position and velocity of each particle in the particle swarm: , ,in: Representing each particle swarm, there are a total of A swarm of particles.

[0035] 3.2 Calculate and compare the fitness of all particle swarms, specifically: Where: T is the total duration of the control loop discretization time. The actual output value for each control cycle. This represents the desired output value. As the error decreases, the fitness increases, ensuring that the initial weights obtained through forward backpropagation with smaller control errors have higher fitness, thus ensuring that the population evolves in this direction and obtains the optimal / relatively good initial weights.

[0036] 3.3 Based on particle swarm optimization Current location ,speed , and Update the velocity of each particle swarm and location Specifically: , ,in: , These are the individual optimal learning rate and the global optimal learning rate (or acceleration constant), i.e., the cognitive coefficient and the social coefficient. , for The random numbers between these values ​​are used to randomly adjust the learning rate of the individual optimal position and the global optimal position. As the current inertia weight, The current global best fitness The corresponding particle swarm position vector, where the superscript t represents the number of iterations optimized by the PSO algorithm; The individual optimal fitness of each particle during the iteration process. The corresponding position vector.

[0037] The current inertia weights are updated using the inertia weight decay method to ensure that a global search can be performed initially, while also allowing for fine-tuning within a local range near the final optimal solution. The weight expression used in the round of iteration is: ,in: The maximum weighting coefficient is set. The minimum weighting coefficient is set. The number of iterations optimized for the current PSO algorithm. This represents the total number of PSO iterations.

[0038] 3.4 If the fitness does not meet the expected value, return to step 2) to iterate; otherwise, stop and obtain the final global optimal fitness. Corresponding particle swarm position vector These are the optimal initial weight parameters for the BP network.

[0039] Step 4: Perform forward computation of the BP neural network and output the current PID parameter K. p K i K d And update the PID controller to obtain the control torque, specifically including:

[0040] 4.1 In the k-th discrete time step, the input of each neuron node in the input layer is: ,in: , This indicates the node number of each neuron in the input layer.

[0041] 4.2 The input to each neuron node in the hidden layer is obtained by weighted summation of the input layer and its corresponding weights. Then, it is activated by a positive-negative symmetric sigmoid function. Obtain the output of the neuron node Specifically: , ,in: , This indicates the node number of each hidden layer neuron.

[0042] 4.3 The input to each neuron node in the output layer is obtained by weighted summation of the hidden layer and its corresponding weights. Then, it is activated by a non-negative Sigmoid function. Obtain the output of the neuron node Specifically: , ,in: , This indicates the node number of each neuron in the output layer.

[0043] 4.4 The output values ​​of the three neurons, i.e., the three updated PID parameters, are used to construct the control quantity of the discrete PID controller, specifically: control torque. ,in: , , This means that the output layer contains three neurons. The corresponding update parameters of the PID controller .

[0044] Step 5: Based on the performance index function E(k), perform backpropagation weight updates for the BP neural network using gradient descent. This involves adjusting the negative gradient direction of the selected performance index function with respect to the weighting coefficients, propagating the error back from the output layer, and updating the weight values ​​between network layers layer by layer. Specifically: , Where: the performance index function at the kth discrete time. , For network learning rate, is the inertia coefficient, m represents the hidden layer node number, and n represents the output layer node number.

[0045] The layer-by-layer update preferably includes: a weight update formula between the hidden layer and the output layer. And the weight update formula between the input layer and the hidden layer. ,in: , , l represents the input layer node number, m represents the hidden layer node number, and n represents the output layer node number.

[0046] Step 6: The control torque obtained in Step 4 is used to adjust the momentum wheel speed through the momentum wheel model, thereby controlling the attitude of the cube star. Specifically: ,in: These represent the roll, pitch, and yaw axes of the cubesat. Indicates the current control step. This represents the rate of change of the momentum wheel speed corresponding to the current control cycle. This represents the current weekly momentum wheel speed. This is the input torque command to the momentum wheel. For the moment of inertia of the corresponding axis momentum wheel, The time step of the discretized control cycle.

[0047] Step 7: After completing the wheel speed adjustment control, the attitude sensor outputs the current control step cuboidal attitude and converts it into Euler angles. If the error between the current control step cuboidal attitude and the desired attitude does not meet the task requirements, return to step 4 for iteration until the requirements are met.

[0048] Based on practical application scenario experiments, the lunar probe satellite "ICUBE-Q", which was successfully launched at 17:27 Beijing time on May 3, 2024 and has been successfully carrying out its on-orbit mission, was selected as the CubeSat model for simulation (i.e., the basic parameters shown in Table 1). The above-mentioned CubeSat attitude stabilization control based on PID controller parameter tuning was carried out. The mission requirement was to achieve three-axis attitude stabilization and Earth orientation after the CubeSat attitude was disturbed, that is, the attitude angles converged to 0 in the CubeSat's center of mass orbit coordinate system.

[0049] Table 1 Simulation parameters for cubesat attitude control

[0050] Prior art compared to this invention includes: EM-PID, a traditional PID controller that uses manually adjusted PID parameters; BP-PID, a PID controller that uses a BP neural network for dynamic parameter tuning, with the initial weights of the BP network being randomly fixed; and PSO-BP-PID, a BP-PID controller that uses the inertial weight decay (PSO) algorithm to optimize the initial weights of the BP network, i.e., the algorithm of this invention. The selected comparison indicators are: control overshoot, 2% attitude angle error settling time, and attitude angle steady-state error.

[0051] To make the comparison more convincing, the disturbance scenarios faced by different controllers were kept the same. Even though the cubesat was subjected to the same initial attitude disturbance as shown in Table 2, the same basic structure of the PID controller and the BP-PID control framework were selected. Figure 2 As shown, the internal structure of the same BP neural network and the forward and backward propagation mechanisms are as follows: Figure 3 As shown, and with the initial network learning rate fixed. and initial inertia coefficient .

[0052] Table 2 Initial attitude perturbation values ​​of CubeSats

[0053] When performing the above-mentioned cubesat attitude stabilization control based on PID controller parameter tuning, the optimal learning rate for each parameter is selected. Global optimal position learning rate Range of inertial weight variation .

[0054] Scene 1

[0055] A simulation was performed on a CubeSat subjected to a single small-angle external attitude perturbation. Three different controllers were used to control this simulation scenario, and the resulting Euler angle response diagrams are shown below. Figures 4-6 As shown, the comparative data of its control indicators in the roll, yaw and pitch directions are shown in Tables 3-5.

[0056] Table 3. Control performance of different controllers against a single small-angle attitude disturbance (roll axis)

[0057] Table 4. Control performance of different controllers against a single small-angle attitude disturbance (yaw axis)

[0058] Table 5. Control performance of different controllers against a single small-angle attitude disturbance (pitch axis)

[0059] Analyzing and comparing the control indicators of the three PID controllers in the table, compared with the EM-PID controller, the overshoot of the three axes decreased by an average of 40.53%, the settling time decreased by an average of 36.96%, and the steady-state error decreased by an average of 69.21% under BP-PID control; under PSO-BP-PID control, the overshoot of the three axes decreased by an average of 47.11%, the settling time decreased by an average of 63.30%, and the steady-state error decreased by an average of 89.84%.

[0060] When encountering attitude disturbances, using a BPNN-based parameter self-tuning PID controller can effectively shorten the fluctuations and adjustment time during attitude adjustment, while approximating the target attitude with smaller errors. A BP-PID controller that optimizes the initial weight values ​​using the inertial weight decay (PSO) algorithm shows better performance in control indicators, further reducing parameter tuning and settling time, and converging to the target attitude more accurately.

[0061] Scene 2

[0062] Due to the complex and variable external disturbance moments affecting satellite attitude in the space environment, it is highly likely that the satellite will be superimposed with other disturbances during the control process. For example, attitude disturbances are constantly generated during the deployment of solar panels. This embodiment simulates this scenario by superimposing random disturbance attitude angles and angular velocities at a certain moment during the initial disturbance control process to simulate the attitude change caused by the CubeSat being subjected to other disturbances before it has completed attitude stabilization control. The parameter settings for simulating the scenario of sudden disturbance are shown in Table 6. At 28 simulation time seconds, a large arbitrary attitude disturbance angle and angular velocity are added to the three axes.

[0063] Table 6. Parameters of secondary disturbances encountered during CubeSat attitude control.

[0064] like Figure 7 As shown, the control responses of three PID controllers are presented, and the comparison data of their control indices in the roll, yaw, and pitch directions are shown in Tables 7-9.

[0065] Table 7. Control performance of different controllers in the face of sudden secondary attitude disturbances (roll axis)

[0066] Table 8. Control performance of different controllers against sudden secondary attitude disturbances (yaw axis)

[0067] Table 9. Control performance of different controllers against sudden secondary attitude disturbances (pitch axis)

[0068] Analyzing the control performance of the three PID controllers in the comparison table, the BP-PID controller performed worse than the EM-PID controller, with an average increase of 3.34% in overshoot across the three axes, an average increase of 49.52% in settling time, and an average increase of 64.04% in steady-state error. The PSO-BP-PID controller, on the other hand, still improved the control performance, with an average decrease of 5.75% in overshoot across the three axes, an average decrease of 38.45% in settling time, and an average decrease of 87.59% in steady-state error.

[0069] When encountering sudden secondary disturbances during attitude adjustment, the EM-PID controller with fixed PID parameters cannot adjust the PID parameters in time to cope with the sudden disturbances, resulting in a significant increase in settling time. Simultaneously, because the BP-PID controller cannot adjust the appropriate initial network weights according to random disturbances, it may fail to optimize the control effect, leading to increased overshoot, prolonged settling time, and consequently, increased steady-state error. However, using the PSO-BP-PID parameter tuning method proposed in this invention, the controller with optimized weights still performs excellently in this scenario, even surpassing its performance in the simpler scenario 1. It not only shortens the settling time but also significantly reduces the average overshoot and steady-state error. This controller ensures the optimization of the control process under different scenarios, demonstrating its good robustness.

[0070] Compared with existing technologies, this invention initializes the BP neural network parameters using a weighted decay particle swarm optimization (PSO) algorithm and dynamically adjusts the PID parameters during cubit star stabilization control. This results in better performance in stabilization control after the cubit star is subjected to attitude disturbances, further shortens the attitude stabilization adjustment time, and converges to the target attitude with a smaller steady-state error. It also provides better scenario adaptability and robustness for cubit star attitude stabilization PID controllers in the face of more complex disturbances.

[0071] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A cubesat attitude stabilization method based on PID controller parameter tuning, characterized in that, include: Step 1: Based on the known parameters of the CubeSat model and its attitude control system, establish the mathematical expression of its attitude dynamics model based on the attitude expression method of quaternions; Step 2: Simulate the closed-loop circuit of the cubesat attitude control system; Step 3: Determine the simulation input values ​​and use the PSO optimization algorithm to assign initial weights to each layer of the BP neural network; Step 4: Perform forward calculations using the BP neural network and update the PID controller to obtain the control torque; Step 5: Update the backpropagation weights of the BP neural network using the gradient descent method; Step 6: The required control torque obtained in Step 4 is used to adjust the speed of the momentum wheel through the momentum wheel model to control the attitude of the cube star. After the speed adjustment control is completed, the attitude sensor outputs the current control step cube star attitude. Step 7: If the error between the current control step cubesat attitude and the desired attitude does not meet the task requirements, return to step 4 for iteration until the requirements are met.

2. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 1, characterized in that, The PID controller outputs the desired control quantity from the current state quantity, and the cubic star momentum wheel, modeled in a three-axis orthogonal manner, independently outputs the torques of the roll, yaw and pitch axes by adjusting each one.

3. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 1, characterized in that, The BP neural network comprises: an input layer with three neurons, a hidden layer with seven neurons, and an output layer with three neurons. The neurons in the input layer represent the actual cube star Euler angle value y(k), the reference Euler angle value r(k), and the error e(k) between them, expressed in Euler angle form. The neuron values ​​in the output layer correspond to the three parameters K of the PID controller. p K i K d .

4. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 1, characterized in that, The momentum wheel actuator is a three-axis orthogonally mounted zero-momentum reaction wheel, used to receive three-axis torque commands from the PID controller and independently adjust the speed of each momentum wheel according to the dynamic model, thereby generating the required control torque. Its core function is to provide a standard torque output and execution interface for the control algorithm of this invention. The attitude sensor is used to simulate the measurement characteristics of a real sensor. Its output consists of the real Euler angles calculated by the cube star attitude dynamics model and the superposition of preset white noise or Gaussian noise. This modeling method aims to verify the robustness of the control algorithm of the present invention in the presence of measurement noise. Its core function is to provide the control loop with attitude feedback signals containing typical errors.

5. The cubic star attitude stabilization method based on PID controller parameter tuning according to any one of claims 1-3, characterized in that, Step 3 specifically includes: 3.1 Set the total number of particle swarms n; based on the number of network weight parameters D of the BP neural network, treat each set of initialized weight parameters as a particle swarm, and use D-dimensional vectors respectively. and D-dimensional vectors This represents the position and velocity of each particle in the particle swarm: , ,in: Representing each particle swarm, there are a total of A swarm of particles; 3.2 Calculate and compare the fitness of all particle swarms, specifically: Where: T is the total duration of the control loop discretization time. The actual output value for each control cycle. As the expected output value decreases, the fitness increases. This ensures that the initial weights obtained through forward BP calculation with smaller control errors have higher fitness, thus ensuring that the population evolves in this direction and obtains the optimal / relatively good initial weights. 3.3 Based on particle swarm optimization Current location ,speed , and Update the velocity of each particle swarm and location Specifically: , ,in: , These are the individual optimal learning rate and the global optimal learning rate (or acceleration constant), i.e., the cognitive coefficient and the social coefficient. , for The random numbers between these values ​​are used to randomly adjust the learning rate of the individual optimal position and the global optimal position. As the current inertia weight, The current global best fitness The corresponding particle swarm position vector, where the superscript t represents the number of iterations optimized by the PSO algorithm; The individual optimal fitness of each particle during the iteration process. The corresponding position vector; 3.4 If the fitness does not meet the expected value, return to step 2) to iterate; otherwise, stop and obtain the final global optimal fitness. Corresponding particle swarm position vector These are the optimal initial weight parameters for the BP network.

6. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 5, characterized in that, The current inertia weights are updated using the inertia weight decay method to ensure that a global search can be performed initially, while also allowing for fine-tuning within a local range near the final optimal solution. The weight expression used in the round of iteration is: ,in: The maximum weighting coefficient is set. The minimum weighting coefficient is set. The number of iterations optimized for the current PSO algorithm. This represents the total number of PSO iterations.

7. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 5, characterized in that, Step 4 specifically includes: 4.1 In the k-th discrete time step, the input of each neuron node in the input layer is: ,in: , Indicates the node number of each input layer neuron; 4.2 The input to each neuron node in the hidden layer is obtained by weighted summation of the input layer and its corresponding weights. Then, it is activated by a positive-negative symmetric sigmoid function. Obtain the output of the neuron node Specifically: , ,in: , Indicates the node number of each hidden layer neuron; 4.3 The input to each neuron node in the output layer is obtained by weighted summation of the hidden layer and its corresponding weights. Then, it is activated by a non-negative Sigmoid function. Obtain the output of the neuron node Specifically: , ,in: , Indicates the node number of each neuron in the output layer; 4.4 The output values ​​of the three neurons, i.e., the three updated PID parameters, are used to construct the control quantity of the discrete PID controller, specifically: control torque. ,in: , , This represents the three neurons in the output layer. The corresponding update parameters of the PID controller .

8. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 5, characterized in that, Step 5, namely, updating the backpropagation weights of the BP neural network using gradient descent based on the performance index function E(k), involves adjusting the negative gradient direction of the selected performance index function with respect to the weighting coefficients, propagating the error back from the output layer, and updating the weight values ​​between network layers layer by layer. ， Where: the performance index function at the kth discrete time. , For network learning rate, Here, m represents the hidden layer node number, and n represents the output layer node number; The layer-by-layer update includes: the weight update formula between the hidden layer and the output layer. And the weight update formula between the input layer and the hidden layer. ,in: , , l represents the input layer node number, m represents the hidden layer node number, and n represents the output layer node number.

9. The cubic star attitude stabilization method based on PID controller parameter tuning according to claim 8, characterized in that, Step 6, namely the control torque obtained in step 4, is used to adjust the momentum wheel speed through the momentum wheel model to control the attitude of the cube star. Specifically: ,in: These represent the roll, pitch, and yaw axes of the cubesat. Indicates the current control step. This represents the rate of change of the momentum wheel speed corresponding to the current control cycle. The current weekly momentum wheel speed, This is the input torque command to the momentum wheel. For the moment of inertia of the corresponding axis momentum wheel, The time step of the discretized control cycle.

10. A cubesat attitude stabilization system based on PID controller parameter tuning for implementing the method of any one of claims 1-9, characterized in that, include: The system includes an initialization parameter optimization module, an attitude determination module, a control law calculation module, and an actuator drive module. The initialization parameter optimization module performs initial optimization of neural network weights based on inertial weight decay (PSO) and online backpropagation weight update processing based on the initial state of the system and the attitude error performance index information during the control process, so as to obtain the BP neural network weight parameter results that optimize the system performance. The attitude determination module calculates the attitude error and performs signal preprocessing based on the noisy Euler angle data measured by the attitude sensor and the desired attitude command information, obtaining the Euler angle error vector and angular velocity information results of the current attitude and the desired attitude. The control law calculation module performs forward calculation of BP neural network and PID control law calculation based on the attitude error information output by the attitude determination module and the initial weight information provided by the initialization parameter optimization module, obtaining the three-axis torque command results for attitude control. The actuator drive module consists of three orthogonally mounted momentum wheels. Based on the three-axis torque command information output by the control law calculation module, it performs torque distribution to obtain the speed adjustment command results of the three zero momentum wheels.

Images