Knowledge and data dual-driven vibration suppression control method for space microgravity vibration isolation system

Abstract

The present invention relates to the field of space technology and vibration control, and in particular to a vibration suppression control method for a space microgravity vibration isolation system driven by both knowledge and data. First, a dynamic model of a space microgravity vibration isolation system is constructed and used as an initial knowledge-driven module. Then, a knowledge-driven nominal preset performance control law is designed; based on the constructed vibration isolation dynamics model, a preliminary vibration suppression control strategy is designed. Secondly, a data-driven vibration suppression strategy and update law are designed. Through the data-driven module and real-time monitored vibration data, a neural network algorithm is used to approximate the uncertainty and disturbance conditions in the current environment. Furthermore, sample generation and network training based on the nominal controller are carried out to simulate the system dynamics behavior under different disturbances. Finally, online training and application based on dual drive of knowledge and data are performed. The present invention can perform adaptive adjustments for different vibration sources and provide efficient and accurate vibration isolation control.

Description

Technical Field

[0001] The present invention relates to the field of space technology and vibration control, and in particular to a knowledge and data dual-driven vibration suppression control method for a space microgravity vibration isolation system. Background Art

[0002] With the rapid development of space technology, vibration isolation control in space environment plays a vital role in microgravity experiments, precision instrument operation and stability assurance of spacecraft platforms. The main function of the vibration isolation system is to reduce the interference of external vibration on equipment operation and ensure the stability and optimal performance of the system. However, vibration isolation systems face many challenges in practical applications, especially when dealing with diverse vibration sources. These vibration sources include not only mechanical vibration and environmental disturbances, but also the noise generated inside the equipment. Traditional control methods show obvious limitations in capturing these complex vibration characteristics, especially when faced with the complexity and uncertainty of the external environment, it is difficult to achieve effective vibration suppression control.

[0003] The uniqueness of the microgravity environment further exacerbates these problems. Since it is difficult for experimental environments on Earth to fully simulate the microgravity state, many existing control algorithms are unable to cope with unexpected external disturbances in actual space missions, resulting in unstable system performance. In addition, traditional linear and nonlinear control methods often show insufficient adaptability when faced with uncertain vibration sources, especially when the vibration source changes dynamically, it is difficult to maintain the optimal performance of the control system. In order to overcome these problems, vibration isolation control methods with stronger adaptability have become the current research focus in the field of vibration isolation.

[0004] Wan Lei, Zeng Jiangfeng, Li Yueming, etc., in "An Intelligent Learning Control Method for Tracking of Unmanned Boats" (CN109283842B), adopted a fuzzy adaptive controller based on model online learning and used a product inference engine to realize the reasoning of rule premises and rule conclusions, which can effectively shorten the adjustment time of the unmanned boat to reach the desired path, enhance the robustness of the control system, and reduce the dependence on the precise mathematical model of the unmanned boat, with faster tracking control response speed and smaller tracking error. Although this method achieves faster response speed and reduces dependence on precise mathematical models through a fuzzy adaptive controller, it still has limitations: fuzzy reasoning is more complex, and its adaptive control may not be robust enough in the face of multi-dimensional and multi-variable complex environments. In addition, fuzzy control may require additional parameter adjustment to maintain control accuracy and efficiency in a spatial vibration environment with more complex uncertainties.

[0005] Zhang Xinglong, Lu Yang, Xu Xin, etc. used a data-driven method to abstract the nonlinear characteristics of the robot into a high-dimensional linear space to form a global linearization model in "A data-driven robust learning predictive control method for robots" (CN110647042B). Based on the global linearization model and the state variables in the high-dimensional linear space, rolling optimization is performed to learn the approximate optimal predictive control sequence of the robot system at the current moment. This method converts complex nonlinear systems into processable linear models through data-driven and Koopman operator technology, thereby improving the description accuracy and control effect of the nonlinear characteristics of the system. Although this method converts complex nonlinear systems into processable linear models in a data-driven manner, thereby improving the description accuracy of nonlinear characteristics, this conversion relies on a large amount of historical data for rolling optimization. In real-time applications, the data-driven learning process may cause a lag in model updates due to the large amount of data and computational complexity, making it difficult to cope with rapidly changing vibration sources. In addition, the linearized model may not be able to fully cope with the highly nonlinear and complex disturbances in the space microgravity environment.

[0006] In "Data-driven control method combining feedforward adjustment with iterative learning" (see "Control Theory and Applications", June 2020, Issue 6, pp. 1376-1376), Bo Yumeng, Cao Mingsheng, and Gao Huibin introduced a parameterized feedforward controller and an input shaping filter, and used the gradient descent method to solve the optimal system feedforward controller. The hybrid data-driven control method that combines feedforward adjustment with iterative learning control improves control accuracy and reduces dependence on system models. Although this method improves control accuracy and reduces dependence on system models by combining feedforward control with iterative learning, it still requires certain system parameterization control. For highly dynamic spatial vibration sources, the design of the feedforward controller may not be able to be updated in real time with disturbance changes. At the same time, although the iterative learning method improves adaptability, the iterative convergence speed is still a bottleneck in the face of completely unknown environments.

[0007] In “Data-driven nonlinear approximate optimal regulation based on iterative neural dynamic programming” (see Acta Automatica Sinica, March 2017, No. 3, pp. 366-375), Wang Ding, Mu Chaoxu, and Liu Derong used the adaptive and self-learning characteristics of neural networks to construct three neural networks (model network, judgment network, and execution network), and used heuristic dynamic programming technology to train the execution network so that it does not rely on system dynamic information and achieves the control effect of the approximate optimal control method. This method uses the self-learning characteristics of neural networks for optimal regulation, but the neural network model relies heavily on the amount of data and the training process. Especially when dealing with nonlinear and unknown disturbance environments, the amount of calculation and time required to train the neural network are long. In addition, although this method can achieve approximate optimal control under ideal conditions, the efficiency and convergence speed of adaptive adjustment are still the main limitations when facing the uncertainty of spatial vibration sources. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a vibration suppression control method for a vibration isolation system driven by both knowledge and data. By combining the knowledge model with the data-driven intelligent control algorithm, this method can not only cope with complex nonlinear vibration problems, but also dynamically adjust according to environmental changes during real-time operation to improve the robustness and response speed of the system. Although traditional adaptive control methods can improve system performance by adjusting control parameters online, they usually rely on predefined model structures and are difficult to handle complex nonlinear systems and uncertain vibration environments. By combining knowledge-driven with data-driven, the present invention can optimize the parameters of the controller through offline multi-task learning and realize real-time and efficient vibration suppression control during operation.

[0009] The technical solution of the present invention:

[0010] A knowledge and data dual-driven vibration suppression control method for a space microgravity vibration isolation system comprises the following steps:

[0011] Step 1: Construction of dynamic model of space microgravity vibration isolation system. First, based on historical data and expert experience, a vibration isolation dynamic model in a microgravity environment is constructed. This model takes into account the dynamic characteristics of the space microgravity vibration isolation system and uses it as the initial knowledge-driven module.

[0012] Step 2: Knowledge-driven nominal preset performance control law design. Based on the vibration isolation dynamics model constructed above, a preliminary vibration suppression control strategy is designed. This strategy uses the existing knowledge model to provide a basic control law. This control law can ensure the stability of the system during operation.

[0013] Step 3: Data-driven vibration suppression strategy and update law design. Next, the system enters the data-driven module, using the real-time monitoring vibration data and the neural network algorithm to approximate the uncertainty and disturbance in the current environment.

[0014] Step 4: Sample generation and network training based on the nominal controller. In this step, the system randomly generates There are corresponding trajectories to simulate the system dynamics behavior under different disturbances. Each trajectory contains vibration data, system response and control law changes within n seconds. By fitting and training the data of the trajectory, the neural network model can approximate the dynamic model under different disturbance conditions and gradually learn how to dynamically adjust according to the changing vibration source. This step significantly enhances the robustness and adaptability of the system's vibration suppression control strategy.

[0015] Step 5: Online training and application based on knowledge and data dual drive. During the actual operation of the system, the online training phase begins. At this point, the data-driven module uses the previous training results as the initial value, and through rapid adjustment and update, it can quickly converge to the optimal control parameters that adapt to the current vibration source. In a dynamic and complex vibration environment, the data-driven module can adaptively adjust the control strategy based on real-time feedback to ensure that the system always maintains the best vibration isolation effect.

[0016] Beneficial effects of the present invention:

[0017] 1. By combining the knowledge-driven and data-driven dual-drive control methods, the system can respond quickly at startup, and use the dynamic model based on historical data and expert experience to provide a preliminary control strategy to ensure that the system can effectively converge and suppress vibration during the startup and operation phases in a microgravity environment.

[0018] 2. Through the data-driven module, the vibration source is monitored and learned in real time by the neural network algorithm, and it can be dynamically adjusted according to different disturbances in the environment. The data-driven algorithm adapts to the changing vibration source through online training, achieves efficient adaptation to complex and dynamic environments, and significantly improves the robustness and adaptability of the system.

[0019] 3. Optimize the neural network parameters under different simulation scenarios, and generate a set of initial values ​​even when the vibration source cannot be predicted in advance, and converge quickly without retraining. In contrast, the direct data-driven control method needs to be retrained when facing new tasks or environmental changes, and has poor adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1It is a flow chart of the space microgravity vibration isolation control method driven by both knowledge and data;

[0021] Figure 2 is the position trajectory map;

[0022] Figure 3 is the posture trajectory diagram;

[0023] Figure 4 is the acceleration trajectory diagram;

[0024] Figure 5 It is the result of the convergence of neural network weights. DETAILED DESCRIPTION

[0025] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0026] like Figure 1 As shown, a knowledge and data dual-driven vibration suppression control method for a space microgravity vibration isolation system of the present invention comprises the following steps:

[0027] Step one: Construction of dynamic model of space microgravity vibration isolation system.

[0028] Based on historical data and expert experience, a dynamic model of a six-degree-of-freedom microgravity vibration isolation system with disturbance terms in a microgravity environment is constructed. The model takes into account the dynamic characteristics of the space microgravity vibration isolation system and uses it as the initial knowledge-driven module.

[0029] The established dynamic model of the six-degree-of-freedom microgravity vibration isolation system with disturbance terms is as follows:

[0030] (1)

[0031] in, is the state variable matrix, for The derivative of is the dynamics simulation matrix, is the control matrix, Indicates the control quantity input, represents the interference matrix. Where: X = [ x 1 x 2 x 3 x 4 x 5 ] T , , , , and They are the state variables of the system: displacement, velocity, acceleration, angle and angular velocity. f = [ f 1 f 2 f 3 f 4 f 5 ] T , , , , , , where represents the natural frequency, Indicates quality, represents the moment of inertia, represents the stiffness matrix, represents the damping matrix, represents the stiffness matrix, Shows the damping matrix. G = [ 0 3 × 3 0 3 × 3 g 3 0 3 × 3 0 3 × 3 0 3 × 3 0 3 × 3 0 3 × 3 0 3 × 3 g 5 ] T , represents the zero matrix, represents the identity matrix, , . u = [ u 3 , u 5 ] T , u 3 = [ u 31 u 32 u 33 ] T represents the acceleration control input, where , , Respectively Acceleration control input for three degrees of freedom, u 5 = [ u 51 u 52 u 53 ] T represents the angular velocity control input, where , , Respectively Angular velocity control input for three degrees of freedom. d = [ 0 3 × 1 , 0 3 × 1 , d 3 , 0 3 × 1 , d 5 ] T , where represents the zero vector, represents the acceleration disturbance, represents the angular velocity disturbance.

[0032] Step 2: Knowledge-driven design of control law with nominal preset performance.

[0033] Based on the vibration isolation dynamics model constructed in the first step, a preliminary vibration suppression control strategy is designed. The preset performance function with flexible design boundaries is introduced, and the input saturation problem common in actual control systems is effectively handled. By adaptively adjusting the performance boundary, the problem of tracking error contacting or exceeding the performance constraint boundary caused by input saturation in traditional methods is avoided. This strategy uses the existing knowledge model to provide a basic control law. This control law can ensure the stability of the system during operation.

[0034] Define the error variable as follows:

[0035] (2)

[0036] in, , , , and are the errors of the system’s state variables displacement, velocity, acceleration, angle, and angular velocity, respectively. , , Respectively represent the filtered , , The virtual control law is obtained through the filter:

[0037] (3)

[0038] in, Represents the virtual control quantity vector of the system, defined as x i − 1 , d = [ x ( i − 1 ) 1 , d x ( i − 1 ) 2 , d x ( i − 1 ) 3 , d ] T , in the vector They represent virtual control quantities with different degrees of freedom respectively; for The derivative of represents the filter constant, Represents the state quantity corresponding to the system. The control law is designed as follows for each degree of freedom:

[0039] (4)

[0040] In the formula, , , , and represents a constant greater than zero, , , , and They represent the error components in each degree of freedom, , , , and Represent the dynamic matrix , , , and The component of each degree of freedom in each state, , and represents the derivative of the virtual control quantity, and Respectively and The component in each degree of freedom.

[0041] Step 3: Design of update law for data-driven vibration suppression strategy.

[0042] Enter the data-driven module, and use the neural network algorithm to approximate the uncertainty and disturbance in the current environment through real-time monitoring of vibration data.

[0043] Define RBF neural network approximation disturbance and :

[0044] (5)

[0045] in, and represents the weight matrix of the neural network, and represents the transpose of the weight matrix of the neural network, and represents the radial basis function, and Represents the approximation error. The estimated value of the error and It is expressed as:

[0046] (6)

[0047] in, and Represents the transpose of the estimated value of the weight matrix of the neural network. The estimated error of the weight matrix is ​​defined as:

[0048] (7)

[0049] The error update law of the weight matrix is ​​designed as:

[0050] (8)

[0051] in, and represents the derivative of the weight matrix, and Represents the weight matrix update gain.

[0052] According to formula (6), the data-driven control law is obtained, and its component design in each degree of freedom is:

[0053] (9)

[0054] In the formula, and They represent the estimated values ​​of the error and Components in three degrees of freedom.

[0055] Step 4: Sample generation and network training based on the nominal controller.

[0056] Through the designed control law and dynamic model, random generation There are corresponding trajectories to simulate the system dynamics behavior under different disturbances. The initial value of the state quantity of each trajectory is different, and the disturbance is different. Each trajectory contains the vibration data, system response and control law changes within n seconds. By fitting and training the data of the trajectory, the neural network model can approximate the dynamic model under different disturbance conditions and gradually learn how to dynamically adjust according to the changing vibration source. This step significantly enhances the robustness and adaptability of the system's vibration suppression control strategy.

[0057] The neural network is trained for each trajectory to obtain the convergence value of the weight matrix. Finally, the initial weight of the neural network as the data-driven module is defined as:

[0058] (10)

[0059] in, and They represent the initial weights used by the data-driven module to train the neural network, and It represents the weight matrix of the reference trajectory used for training after training.

[0060] Step 5: Online training and application driven by both knowledge and data.

[0061] Based on the offline training results and real-time monitoring data of the vibration source, the data-driven module uses the previous training results as the initial value, and dynamically approximates the disturbance through the real-time collected vibration data combined with the neural network.

[0062] The sensor is used to monitor the vibration of the system in real time and collect data of the current vibration source, including vibration amplitude, frequency and corresponding disturbance signal. The neural network is updated online by using the collected vibration data and the neural network model shown in formula (10) as the initial value of the neural network. The neural network update law is shown in formula (8). The updated neural network is finally used to estimate the actual disturbance, and combined with the data-driven control law shown in formula (9), the system control strategy is adjusted and optimized through a fast feedback mechanism with real-time data.

[0063] Under different vibration sources and environments, the control parameters are quickly converged through adaptive adjustment and rapid iteration. The adjusted control parameters can adapt to new vibration sources within seconds, ensuring the stable operation of the system in complex and dynamic environments.

[0064] In order to verify the effectiveness of the proposed control technology, the following system parameters are selected for simulation verification of the microgravity vibration isolation system. for:

[0065] J s = [ 0 . 9 2 − 0 . 008 − 0 . 0 4 1 − 0 . 008 1 . 2 0 . 0 0 2 − 0 . 0 4 1 0 . 0 0 2 1 . 5 ] k g ⋅ m 2 (10)

[0066] External disturbance torque from the environment

[0067] d 5 = [ 1 + 2 c o s ( 0 . 3 t ) 2 + 1 . 5 s i n ( 0 . 3 t ) + 3 c o s ( 0 . 3 t ) 3 + 2 s i n ( 0 . 2 t ) ] × 1 0 − 4 N m (11)

[0068] The initial value of the system is [ 5 − 4 . 5 − 1 . 8 ] T , the initial posture angle is [ 5 − 5 5 ] T deg. The desired position of the vibration isolation system is set to [ 0 0 0 ] T The expected angle is [ 0 0 0 ] T During the initial training, based on the previously constructed vibration isolation dynamics model, the system quickly calculates and outputs a preliminary control strategy to deal with the initial vibration disturbance. Then, based on this control strategy, 50 random trajectories are generated, each containing dynamic data (vibration amplitude, frequency, control law, etc.) under different disturbances within 30 seconds.

[0069] The neural network is used to fit these trajectories so that the data-driven model can approach the actual vibration environment and dynamically adjust the control strategy. The sampling frequency is 0.1s. The interference force is a vibration interference combination in the range of 0.01Hz-50Hz, and the interference torque is set as shown in formula (11). The controller parameters are set as follows:

[0070] k 1 = [ 0 . 5 , 0 . 4 , 0 . 3 ] k 2 = [ 1 . 5 , 0 . 8 , 0 . 6 ] k 3 = [ 10 , 10 , 10 ] k 4 = [ 0 . 1 , 0 . 1 , 0 . 1 ] k 5 = [ 1 . 2 , 0 . 8 , 0 . 8 ]

[0071] The simulation curve obtained by optimizing the control parameters after training is as follows: Figures 2 to 5 shown.

[0072] Figure 2 It shows that the initial vibration amplitude was large in the position axis of the system, and the system experienced significant displacement changes. Through the intelligent control strategy, the system gradually suppressed the vibration and controlled the vibration amplitude in the position axis to a stable state within 20 seconds. This result shows that the system can not only effectively reduce large displacements, but also quickly adjust to the desired position to ensure the spatial positioning accuracy of the system.

[0073] Figure 3The system shows obvious vibration in the initial attitude angle, especially in the startup phase, when the angle of the attitude axis changes dramatically. However, with the application of the control strategy, the attitude vibration is gradually suppressed, and the system can adjust the attitude to stability within 20 seconds. This result verifies the system's attitude control capability on multi-degree-of-freedom rotation axes, which can maintain high attitude accuracy and stability and effectively cope with attitude disturbances in the space environment.

[0074] Figure 4 The simulation shows that the acceleration vibration of the system is quite severe in the initial stage, especially under the influence of external disturbances, the acceleration vibration frequency is high. However, through the adaptive control of the system, the system suppresses the acceleration vibration to a stable state in less than 20 seconds. This result shows that the system has a strong suppression ability when dealing with high-frequency vibrations, and can effectively reduce the acceleration changes in a short time, ensuring the stability and safety of the system operation.

[0075] Overall, the system has good vibration control effects in all directions and can control the vibration amplitude to a stable state in a relatively short time. Figure 5 The convergence of the weight norms of different layers of the neural network during the training process is shown. The weights converge quickly and are basically stable in about 20 seconds, indicating that the neural network has achieved good convergence in a relatively short period of time. Although the fluctuations of the weights of each layer after convergence are different, they show good convergence overall, proving the effectiveness of the neural network in vibration control. The change process of the weights also reflects the system's ability to adaptively adjust to different vibration sources through a data-driven model.

Claims

1. A knowledge and data dual-driven vibration suppression control method for a space microgravity vibration isolation system, characterized in that: The following steps are involved: Step 1: Construction of dynamic model of space microgravity vibration isolation system; Based on historical data and expert experience, a six-DOF microgravity vibration isolation system dynamics model with disturbance terms in a microgravity environment is constructed and used as the initial knowledge-driven module; The established dynamic model of the six-degree-of-freedom microgravity vibration isolation system with disturbance terms is as follows: (1) in, is the state variable matrix, for The derivative of is the dynamics simulation matrix, is the control matrix, Indicates the control quantity input, represents the interference matrix; where: , , , , and They are the state variables of the system: displacement, velocity, acceleration, angle, and angular velocity; , , , , , , where represents the natural frequency, Indicates quality, represents the moment of inertia, represents the stiffness matrix, represents the damping matrix, represents the stiffness matrix, Shows the damping matrix; , represents the zero matrix, represents the identity matrix, , ; , represents the acceleration control input, where , , Respectively Acceleration control input for three degrees of freedom, represents the angular velocity control input, where , , Respectively Angular velocity control input for three degrees of freedom; , where represents the zero vector, represents the acceleration disturbance, represents the angular velocity disturbance; Step 2: Knowledge-driven nominal preset performance control law design; Based on the vibration isolation dynamics model constructed in the first step, a preliminary vibration suppression control strategy is designed with a preset performance function with flexible design boundaries, as follows: Define the error variable as follows: (2) in, , , , and are the errors of the system’s state variables displacement, velocity, acceleration, angle, and angular velocity, respectively. , , Respectively represent the filtered , , The virtual control law is obtained through the filter: (3) in, Represents the virtual control quantity vector of the system, defined as , in the vector They represent virtual control quantities with different degrees of freedom respectively; for The derivative of represents the filter constant, Represents the state quantity corresponding to the system; the control law component in each degree of freedom is designed as: (4) In the formula, , , , and represents a constant greater than zero, , , , and They represent the error components in each degree of freedom, , , , and Represent the dynamic matrix , , , and The component of each degree of freedom in each state, , and represents the derivative of the virtual control quantity, and Respectively and The component in each degree of freedom; Step 3: Design of update law for data-driven vibration suppression strategy; Enter the data-driven module, and use the neural network algorithm to approximate the uncertainty and disturbance in the current environment through real-time monitoring of vibration data; Define RBF neural network approximation disturbance and : (5) in, and represents the weight matrix of the neural network, and represents the transpose of the weight matrix of the neural network, and represents the radial basis function, and represents the approximation error; the estimated value of the error and It is expressed as: (6) in, and Represents the transpose of the estimated value of the weight matrix of the neural network; the estimated error of the weight matrix is ​​defined as: (7) The error update law of the weight matrix is ​​designed as: (8) in, and represents the derivative of the weight matrix, and represents the weight matrix update gain; According to formula (6), the data-driven control law is obtained, and its component design in each degree of freedom is: (9) In the formula, and They represent the estimated values ​​of the error and Components in three degrees of freedom; Step 4: Sample generation and network training based on the nominal controller; Through the designed control law and dynamic model, random generation corresponding trajectories to simulate the system dynamics behavior under different disturbances; By fitting and training the data of the trajectories, the neural network model can approximate the dynamic model under different disturbance conditions and gradually learn how to dynamically adjust according to the changing vibration source; The neural network is trained for each trajectory to obtain the convergence value of the weight matrix. Finally, the initial weight of the neural network as the data-driven module is defined as: (10) in, and They represent the initial weights used by the data-driven module to train the neural network, and Then represents the weight matrix of the reference trajectory used for training after training convergence; Step 5: Online training and application based on knowledge and data dual drive; Based on the offline training results and the real-time monitoring data of the vibration source, the data-driven module uses the previous training results as the initial value, and dynamically approximates the disturbance through the real-time collected vibration data combined with the neural network; the details are as follows: The vibration of the system is monitored in real time by using sensors to collect data of the current vibration source, including vibration amplitude, frequency and corresponding disturbance signals. The neural network model shown in formula (10) is used as the initial value of the neural network through the collected vibration data to update the neural network online. The neural network update law is shown in formula (8). Finally, the updated neural network is used to estimate the actual disturbance, and combined with the data-driven control law shown in formula (9), the system control strategy is adjusted and optimized through a fast feedback mechanism with real-time data.

Images