On-orbit learning method and device for takeover control of non-cooperative spacecraft

Abstract

The invention discloses an on-orbit learning method and device for non-cooperative spacecraft takeover control. According to the method, the on-orbit learning method is designed by using the neural network, and the unknown torque of the combined spacecraft system is predicted and compensated, so that stable control can be realized in a more complex and changeable power environment, and the application scene is wider than that of the existing method. Meanwhile, the on-orbit learning method can perform deep learning and mining on historical on-orbit data, thereby effectively coping with relatively large uncertainty in the system, and remarkably improving the flexibility and robustness of the control algorithm.

Description

Technical Field

[0001] This invention relates to the field of spacecraft dynamics and control technology, specifically to an on-orbit learning method and apparatus for takeover control of non-cooperative spacecraft. Background Technology

[0002] Spacecraft attitude takeover control typically involves a service spacecraft coupling with a target spacecraft to form a combined spacecraft, with the service spacecraft's propulsion system controlling the combined spacecraft's attitude adjustments. There are two main coupling methods: one utilizes a pre-defined interface between a space robotic arm and the target spacecraft, and the other involves a small spacecraft attaching itself to the target spacecraft's surface to achieve coupling. Non-cooperative spacecraft refer to space targets in orbit that lack the ability to actively communicate, coordinate navigation, or cooperate with the service spacecraft. They typically do not possess a specific interface for cooperation with a space robotic arm, making it difficult to achieve takeover tasks using a space robotic arm in this scenario. Small spacecraft can attach to the target spacecraft's surface through SLAM (Simultaneous Localization and Mapping) technology, involving four stages: target search, anchor point discovery, anchor point tracking, and cooperation. [3] It has excellent adaptability to spacecraft of different shapes and sizes, and features low launch cost and short development cycle, making it a promising candidate for attitude control in non-cooperative spacecraft scenarios.

[0003] In existing technologies, most control methods rely on accurate dynamic system modeling. However, in non-cooperative spacecraft takeover control scenarios, there are often significant unknown disturbances that are difficult to model, making these model-based control methods unsuitable. Furthermore, most control algorithms designed for spacecraft takeover scenarios assume that the target spacecraft has lost its active torque output capability. However, in actual missions, non-cooperative spacecraft often still possess active maneuvering capabilities, making attitude takeover control of non-cooperative spacecraft even more challenging.

[0004] In 2021, Zhang, C., Ahn, CK, Wu, J., & He, W. proposed an online-learning control method for spacecraft attitude tracking (Zhang, C., Ahn, CK, Wu, J., & He, W. Online-learning control with weakened saturation response to attitude tracking: A variable learning intensity approach. Aerospace Science and Technology, 117, 106981, 2021). This method incorporates the control torque information from the previous moment into the control law to achieve online adaptive adjustment of controller parameters. A parameter adjustment mechanism is also designed to prevent the control torque from remaining in a saturated state for extended periods. The stability of the system is rigorously proven theoretically. Simulation results show that this method can achieve high-precision and robust control of spacecraft attitude. However, this method is essentially still a model-based control method, and its robustness depends on accurate dynamic modeling. In the attitude takeover control scenario of non-cooperative spacecraft, the system often has large unknown torques, making accurate modeling difficult and hindering the application of existing methods. Furthermore, relying solely on a single frame of historical data is insufficient to learn enough on-orbit information; therefore, the improvement in control performance offered by this method is limited. Summary of the Invention

[0005] In view of this, the present invention provides an on-orbit learning method for takeover control of non-cooperative spacecraft. After offline learning, it can realize real-time prediction of unknown torques of the target spacecraft, and then combine with a compensation mechanism to realize attitude takeover control of non-cooperative spacecraft.

[0006] The on-orbit learning method for takeover control of non-cooperative spacecraft of the present invention employs a takeover module attached to the surface of the target spacecraft to form a combined spacecraft. The attitude control of the combined spacecraft is achieved by the torque output of the takeover module. Specifically, it includes: S1, the attitude dynamics model of the combined spacecraft is constructed as follows: (1) in, The moment of inertia of the combined spacecraft; To determine the angular velocity of the spacecraft's body coordinate system relative to the inertial coordinate system; For the attitude quaternions of the combined spacecraft; The output torque of the target spacecraft; The output torque of the control module; External disturbance torque; operator This represents the cross product matrix generated by the vectors; the superscript T indicates transpose. It is the identity matrix; S2, Construct the neural network; The input to the neural network is... , The quaternion for the attitude error of the target spacecraft. , Desired attitude of the target spacecraft The conjugate quaternion; The output of the neural network is , The predicted value of the output torque for the target spacecraft; S3, the control module applies a certain excitation torque, measures the attitude angle and angular velocity of the combined spacecraft under this excitation torque, and uses... Estimate the output torque of the target spacecraft and construct a sample dataset. ; Based on the sample dataset Train the neural network constructed by S2; S4, output torque of the control module ,in, The trained neural network is used to predict the output torque of the target spacecraft in real time, in order to compensate for and counteract the unknown torque of the system. The torque required for the combined spacecraft to reach the target attitude when there is no unknown torque in the system can be calculated using a common negative feedback control law based on the current state error. Preferably, the neural network employs a multilayer perceptron network, a radial basis function neural network, a recurrent neural network, a long short-term memory periodic network, a convolutional neural network, or a Transformer network.

[0007] Preferably, in S3, the excitation torque is a sinusoidal excitation torque, a square wave excitation torque, a noise excitation torque, a frequency scanning excitation torque, or a pulse excitation torque.

[0008] Preferably, in step S4, the takeover module uses PD control, state feedback control, LQR optimal control, or model reference adaptive control to calculate... .

[0009] The present invention also provides an attitude takeover control device for a non-cooperative spacecraft, comprising: Sensors are used to collect the attitude angles and angular velocities of the combined spacecraft; The memory includes: external memory for storing neural network learning and prediction programs; and internal memory for storing data collected by sensors and trained neural network parameters. The processor, based on sensor data and neural network parameters stored in memory, calculates the output torque of the control module using the method described above. The actuator is used to execute the output torque calculated by the processor to complete the attitude control of the target spacecraft.

[0010] Beneficial effects: This invention utilizes a neural network to design an on-orbit learning method for predicting and compensating for unknown torques in combined spacecraft systems. This enables stable control in more complex and variable dynamic environments, broadening its application scenarios compared to existing methods. Furthermore, this on-orbit learning method can perform deep learning and mining of historical on-orbit data, effectively addressing significant uncertainties in the system and significantly improving the flexibility and robustness of the control algorithm.

[0011] The control device of this invention has a compact structure and small size, making it suitable for mission scenarios where attitude takeover control is performed using tiny spacecraft. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the implementation of the on-orbit learning method of the present invention.

[0013] Figure 2 This refers to the quaternion changes in the online control process.

[0014] Figure 3 This refers to the torque prediction results during the online control process.

[0015] Figure 4 This is a schematic diagram of the control device of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] The torque output of non-cooperative spacecraft is typically calculated based on its attitude angles and angular velocities, indicating a functional relationship between the introduced system disturbances and the real-time state information. This invention addresses the mission characteristics of attitude takeover control of non-cooperative spacecraft by proposing an on-orbit learning method. This method utilizes a neural network to learn from the data collected by the spacecraft in orbit, constructing a mapping model between unknown disturbance torques and system states. This enables real-time prediction of unknown system torques and real-time correction of control commands through an online compensation mechanism, achieving attitude takeover control of the non-cooperative spacecraft with strong adaptability and robustness.

[0018] The flowchart of the method of this invention is as follows Figure 1 As shown, the specific steps include the following: Step 1: Assembled spacecraft attitude dynamics modeling The control method proposed in this invention is achieved through a small spacecraft, referred to here as a control module. The control module is attached to the surface of the target spacecraft to form a combined spacecraft, independently outputting torque to achieve attitude control of the spacecraft. A volume coordinate system is established with the combined spacecraft as the reference frame. Establish an inertial coordinate system with Earth as the reference frame. The attitude of the combined spacecraft can be determined through... Compared to The rotation is used to represent the attitude dynamics, and quaternions are introduced to perform attitude dynamics calculations. The combined spacecraft is controlled by two propulsion systems: the takeover module and the target spacecraft. Therefore, the attitude dynamics model of the combined spacecraft can be modeled as follows: (1) in, The moment of inertia of the combined spacecraft; The combined spacecraft angular velocity is the angular velocity of the combined spacecraft body coordinate system relative to the inertial coordinate system. For the attitude quaternions of the combined spacecraft, and These represent the vector and scalar parts of a quaternion, respectively. The output torque of the target spacecraft; The output torque of the control module; External disturbance torque; operator This represents the cross product matrix generated by the vectors; the superscript T indicates transpose. It is an identity matrix.

[0019] Both control systems of the combined spacecraft aim to achieve their respective desired attitudes. Assume the desired attitude of the target spacecraft is... Then its error quaternion It can be calculated using the following formula (2) Operators This represents quaternion multiplication. According to the rules of quaternion multiplication, we have... (3) Similarly, the desired attitude of the takeover module. and attitude error ,have (4) The control objective of this invention is to achieve the desired result in equation (1). The error quaternion calculated by equation (4) converges to the interval near 0 under the condition that it exists and is unknown, and the system can remain stable.

[0020] Step 2: Spacecraft On-Orbit Learning Method Based on Multilayer Perceptron Step 2.1, Offline Learning Process Taking over control of non-cooperative spacecraft using on-orbit learning methods involves two processes: offline learning and online control. In the offline learning phase, the functional relationship between the combined spacecraft state and unknown torque is reconstructed using data collected on-orbit. In the online control phase, attitude takeover control of the non-cooperative spacecraft is achieved through a compensation mechanism.

[0021] This embodiment achieves on-orbit learning for the spacecraft by training a Multilayer Perceptron (MLP) network. An MLP is a feedforward neural network composed of multiple fully connected neurons, capable of mapping input data to one-dimensional or multi-dimensional output data. Typically, the output of the spacecraft controller is determined by the spacecraft's state information and desired attitude; therefore, the… Modeling as (5) Therefore, the trainable input layer is: Output layer is The MLP network is used to reconstruct the mapping model between spacecraft state and unknown torques, where express The predicted value. Input data The output data used for training can be obtained through sensor measurements of the control module and the solution of equation (3). It can be estimated according to equation (1), and we have (6) in express The estimated value of angular acceleration Through the Obtained by first-order difference. Ignore. The estimation error introduced can be reduced by applying a certain excitation torque to the control module, such as sinusoidal excitation torque, square wave excitation torque, white noise excitation torque, frequency scanning excitation torque, pulsed excitation torque, etc. Data sets can be obtained by collecting data over a period of time at a certain sampling frequency. .

[0022] Adding several hidden layers between the input and output layers, each containing several nodes for high-dimensional mapping of the data from each layer, then the... The output of each hidden layer is (7) in and For the weight matrix and the bias vector, Let represent the activation function. If no activation function is added to the output layer, the MLP network can perform regression tasks. Using the backpropagation algorithm, the mean squared error is chosen as the loss function to train the MLP network, resulting in a network capable of performing regression tasks. Make predictions.

[0023] This invention can also employ radial basis function neural networks, recurrent neural networks, long short-term memory periodic networks, convolutional neural networks, Transformer networks, and other networks. To make predictions, the mean absolute error (MAE) can also be used as the loss function to train the neural network.

[0024] Step 2.2, Online Control Process Will The control law is designed as (8) Substituting into equation (1) (9) Prediction error of neural networks It is bounded, and this will be verified through simulation examples, external interference. It is also generally considered to be bounded, therefore the term It can be regarded as a bounded unknown disturbance, and thus the attitude takeover control problem can be transformed into a general combined spacecraft attitude control problem that does not consider the unknown torque of the target spacecraft through the compensation mechanism of Equation (8). Used to achieve stable control in equation (9). Regarding the control law... Numerous studies have explored various methods for achieving stable attitude control, such as PD control, state feedback control, LQR optimal control, and model reference adaptive control. This invention primarily introduces a control compensation method, which will not be elaborated upon here.

[0025] In summary, the on-orbit learning method proposed in this invention for achieving attitude takeover control of non-cooperative spacecraft consists of two processes: offline learning and online control. During offline learning, the takeover module uses its onboard sensors to measure the combined spacecraft's state in real time. And calculate Data sets are obtained after a period of time. Next, the MLP network is trained using the backpropagation algorithm. After training, the MLP network can be used to perform attitude takeover control of non-cooperative targets. The MLP network uses real-time state information to predict the target spacecraft's output. And use Equation (8) for compensation to achieve stable control of the combined spacecraft.

[0026] Based on the above method, the following simulation scenario is set up: Moment of inertia of combined spacecraft External interference ,in Desired attitude of the target spacecraft ; Takeover module expected posture The target spacecraft control rate is PD control. ,in ; Control rate of the takeover module For PD control ,in .

[0027] Data was collected at a frequency of 10Hz for 2000 seconds for offline training. During this period, the takeover module output a sinusoidal excitation torque. 80% (the first 1600 seconds) of the data was used as the training set, and 20% (the last 400 seconds) was used as the test set. The hidden layers of the MLP network were set to... With a learning rate of 0.001, the trained network is used for real-time prediction and compensation, thereby realizing the simulation of the online control stage. The simulation results are as follows: Figure 2 and Figure 3 As shown. From Figure 2 and Figure 3 It can be seen that the non-cooperative target attitude takeover control method proposed in this invention can achieve the expected control objective.

[0028] Based on the above method, this invention also proposes a control device for implementing the on-orbit learning process, thereby achieving attitude takeover control of non-cooperative spacecraft. The schematic diagram of this device is shown below. Figure 4 As shown, the device uses the state information measured by the sensors of the takeover module as input and outputs the calculated torque to the actuator of the takeover module. The device mainly consists of three parts: external memory, internal memory, and a processor. The external memory is a non-volatile storage medium that stores programs for both offline learning and online control. The internal memory is a volatile storage medium used to store temporary calculation results during program execution, including the collected dataset and trained MLP network parameters, which are input to the external memory for long-term storage. The processor executes the non-cooperative target attitude takeover control program and uses program configuration to achieve logical switching between offline learning and online control programs. Furthermore, the device can be integrated onto a 10cm × 5cm PCB board, ensuring it does not affect the size of the takeover module and enabling its application in tasks requiring non-cooperative target attitude takeover control via the takeover module.

[0029] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An on-orbit learning method for takeover control of non-cooperative spacecraft, comprising a takeover module attached to the surface of a target spacecraft to form a combined spacecraft, wherein the takeover module outputs torque to complete the attitude control of the combined spacecraft, characterized in that, include: S1, the attitude dynamics model of the combined spacecraft is constructed as follows: （1） in, The moment of inertia of the combined spacecraft; To determine the angular velocity of the spacecraft's body coordinate system relative to the inertial coordinate system; For the attitude quaternions of the combined spacecraft; The output torque of the target spacecraft; The output torque of the control module; External disturbance torque; operator This represents the cross product matrix generated by the vectors; the superscript T indicates transpose. It is the identity matrix; S2, Construct the neural network; The input to the neural network is... , The quaternion for the attitude error of the target spacecraft. , Desired attitude of the target spacecraft The conjugate quaternion; The output of the neural network is , The predicted value of the output torque for the target spacecraft; S3, the control module applies a certain excitation torque, measures the attitude angle and angular velocity of the combined spacecraft under this excitation torque, and uses... Estimate the output torque of the target spacecraft and construct a sample dataset. ; Based on the sample dataset Train the neural network constructed by S2; S4, output torque of the control module for: ;in, Real-time prediction of the output torque of the target spacecraft by the neural network trained for S3; This refers to the torque required for the combined spacecraft to achieve the target attitude without considering the output torque of the target spacecraft.

2. The method as described in claim 1, characterized in that, The neural network is constructed using a multilayer perceptron network, a radial basis function neural network, a recurrent neural network, a long short-term memory periodic network, a convolutional neural network, or a Transformer network.

3. The method as described in claim 1 or 2, characterized in that, In S3, during training, the loss function is either mean squared error or mean absolute error.

4. The method as described in claim 1 or 2, characterized in that, In S3, the excitation torque is a sinusoidal excitation torque, a square wave excitation torque, a noise excitation torque, a frequency scanning excitation torque, or a pulse excitation torque.

5. The method as described in claim 1, characterized in that, In step S4, the takeover module uses PD control, state feedback control, LQR optimal control, or model reference adaptive control to calculate... .

6. An attitude control device for a non-cooperative spacecraft, characterized in that, include: Sensors are used to collect the attitude angles and angular velocities of the combined spacecraft; The memory includes: external memory for storing neural network learning and prediction programs; and internal memory for storing data collected by sensors and trained neural network parameters. The processor, based on sensor acquisition data and neural network parameters stored in memory, calculates the output torque of the control module using the method described in any one of claims 1 to 5; The actuator is used to execute the output torque calculated by the processor to complete the attitude control of the target spacecraft.

Images