An intelligent decision-making method for spacecraft attitude and orbit cooperative optimization
By establishing a spacecraft attitude and orbit coupled decision-making model through deep reinforcement learning and training the neural network using asynchronous data generation methods, the problem of traditional methods being difficult to apply in time-sensitive and highly interactive scenarios is solved. This enables collaborative optimization decision-making for spacecraft attitude and orbit, improving the autonomy and adaptability of mission execution.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF CONTROL ENG
- Filing Date
- 2022-09-09
- Publication Date
- 2026-05-29
Smart Images
Figure CN116339360B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an intelligent decision-making method for collaborative optimization of spacecraft attitude and orbit, belonging to the field of spacecraft attitude and orbit control technology. Background Technology
[0002] During on-orbit missions, spacecraft often face real-time scenarios, such as monitoring time-sensitive events, or complex interactions between the spacecraft and its environment, such as close-range operations against non-cooperative targets. In these situations, the spacecraft's attitude and orbit exhibit strong coupling characteristics. If traditional decoupling design approaches are used to determine attitude and orbit strategies, it may negatively impact mission performance or even lead to mission failure. However, due to the dynamic uncertainties of the environment in time-sensitive event or close-range operation scenarios, it is difficult to establish effective analytical models, making traditional model-based attitude and orbit decision-making methods unsuitable. Therefore, for future complex space missions, an intelligent decision-making method that does not rely on analytical models is needed to achieve coordinated decision-making regarding spacecraft attitude and orbit. Summary of the Invention
[0003] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose an intelligent decision-making method for the collaborative optimization of spacecraft attitude and orbit, so as to realize the collaborative decision-making of spacecraft attitude and orbit using deep reinforcement learning methods when the environmental model is unknown.
[0004] The present invention solves the above-mentioned technical problem through the following technical solution:
[0005] An intelligent decision-making method for collaborative optimization of spacecraft attitude and orbit includes:
[0006] Conduct spacecraft orbital dynamics and attitude dynamics simulations, and obtain the spacecraft's position, velocity, attitude, and Euler angular velocity relative to the predetermined mission target with a predetermined simulation iteration step size; use the data obtained above and the relative position of the Sun in the orbital system as a sample pool;
[0007] A spacecraft attitude and orbit coupling decision model is established, including an attitude decision neural network, an orbit decision neural network, and a reward function algorithm for reinforcement learning of the above two decision neural networks. The attitude decision neural network and the orbit decision neural network generate the expected control torque and velocity increment of the spacecraft in the x, y, and z directions under the orbital system, respectively, to control the attitude and orbit of the spacecraft.
[0008] Sample pool data is asynchronously collected at different sampling frequencies and used as training data for attitude decision neural network and orbit decision neural network, respectively. The weights of the two decision neural networks are iteratively updated to complete the network training and obtain a successfully trained spacecraft attitude-orbit coupling decision model.
[0009] The system collects the position, velocity, attitude, Euler angular velocity, and relative position of the spacecraft to be monitored relative to the target, as well as the Sun's position in the orbital system. These data are then fed into a successfully trained spacecraft attitude-orbit coupling decision model to obtain the desired velocity increment and desired torque. This data drives the spacecraft to change its attitude and orbit, achieving an orbital approach to the target.
[0010] Preferably, the neural network for establishing trajectory decision-making includes:
[0011] An orbital decision neural network is established. The input data of the network includes the position and velocity of the spacecraft relative to the predetermined mission target and the relative position of the sun in the orbital system. The output is the expected velocity increment of the spacecraft in the x, y, and z directions in the orbital system.
[0012] Preferably, the neural network for establishing attitude decision-making includes:
[0013] Design an attitude decision neural network. The input data of the network includes the attitude quaternion, attitude angular velocity, position, and velocity of the spacecraft relative to the predetermined mission target. The output is the expected control torque of the spacecraft in the x, y, and z directions in the orbital system.
[0014] Preferably, both the orbital decision neural network and the attitude decision neural network are fully connected neural networks, and the activation function of both networks is the ReLU function.
[0015] Preferably, the reward function algorithm includes a public part and a private part:
[0016] The public part is used to evaluate the completion of the predetermined task, and the same value is used for training the attitude decision neural network and the trajectory decision neural network;
[0017] The private part differs in its training for the attitude decision neural network and the orbit decision neural network: for the orbit decision neural network, the private part is designed as a function related to the velocity increment, while for the attitude decision neural network, the private part is designed as a function related to the control torque.
[0018] Preferably, the reward function algorithm for the orbital decision neural network is as follows:
[0019]
[0020] In the formula, cosθ is the angle between the spacecraft as the vertex, the predetermined mission target, and the sun, and d 12Let |Δv| be the distance between the spacecraft and the predetermined mission target, |Δv| be the sum of the absolute values of the velocity increments in the x, y, and z directions in the orbital system, |ΔT| be the sum of the absolute values of the expected control torques of the spacecraft in the x, y, and z directions, t represent the predetermined mission duration, f represent the predetermined mission completion rate, and t+f be the common part of the reward function.
[0021] The preferred reward function algorithm for the attitude decision neural network is as follows:
[0022] rew 0b2 +=C2+|ΔT|+t+f
[0023] Where C2 is the impact assessment value of the predetermined mission objective on the spacecraft, |ΔT| is the sum of the absolute values of the expected control torques of the spacecraft in the x, y, and z directions, t represents the mission duration, f represents the mission completion rate, and t+f is the common part of the reward function.
[0024] Preferably, the training data for the attitude decision neural network is collected from the sample pool at a time of 20-100 times the simulation iteration step size, and the data for the trajectory decision neural network is collected from the sample pool at a time of 10,000-30,000 times the simulation iteration step size.
[0025] Preferably, during network training, the spacecraft's mass, moment of inertia, and initial state are randomly initialized to improve the generalization ability of the attitude decision neural network and the orbit decision neural network.
[0026] Preferably, the spacecraft's moment of inertia, initial position, initial velocity, and velocity increments in the x, y, and z directions of the orbital system are input into the spacecraft's orbital dynamics model to obtain the spacecraft's position and velocity relative to the predetermined mission target with a predetermined simulation iteration step size; the spacecraft's mass, initial position, initial velocity, and control torques in the x, y, and z directions of the orbital system are input into the spacecraft's attitude dynamics model to obtain the spacecraft's attitude four elements and Euler angular velocity relative to the predetermined mission target with a predetermined simulation iteration step size.
[0027] The advantages of this invention compared to the prior art are:
[0028] (1) This invention provides an intelligent decision-making method for collaborative optimization of spacecraft attitude and orbit. For the first time, it proposes using deep reinforcement learning to make collaborative decisions on spacecraft attitude and orbit under unknown environmental models, solving the problem that traditional design methods are difficult to apply due to the difficulty in establishing accurate analytical models in time-sensitive and highly interactive scenarios. Currently, most spacecraft operate in deterministic environments, mainly considering the relative relationships between the sun, earth, and themselves. Therefore, very accurate relative relationship models can be obtained using orbital dynamics and rigid body attitude dynamics. However, as the complexity of space missions increases, the demand for autonomous spacecraft execution in highly real-time, highly dynamic, and highly interactive scenarios is becoming increasingly urgent. However, due to the difficulty in effectively establishing interaction models between spacecraft and the environment in unknown and highly dynamic environments, traditional model-based design methods are difficult to apply to these new scenarios and tasks. Based on the design concept of proactive exploration and autonomous learning, this invention establishes a collaborative decision-making method for spacecraft attitude and orbit under a deep reinforcement learning framework. It proposes corresponding solutions in training data generation, reward function design, and network structure design, providing a feasible approach for collaborative decision-making of spacecraft attitude and orbit in complex scenarios, given the shortcomings of existing design methods.
[0029] (2) This invention provides an intelligent decision-making method for spacecraft attitude and orbit collaborative optimization, and for the first time proposes an asynchronous data generation method for attitude and orbit, solving the problem that the large difference in attitude and orbit decision cycles affects the efficiency of training data generation. In conventional reinforcement learning scenarios, such as robotic arm control tasks, multiple decision frequencies are generally not involved, and data acquisition is carried out under a unified standard. However, for spacecraft attitude and orbit collaborative decision-making tasks, due to the short-cycle characteristics of attitude and the long-cycle cumulative effect of orbit, if a unified data acquisition standard is still followed, such as sampling according to the attitude decision cycle, then a large number of invalid training samples will exist during the training of the orbit decision network. If sampling is uniformly performed according to the orbit decision cycle, then the attitude decision network will lack training samples. Therefore, this invention proposes an asynchronous data generation method for attitude and orbit, assigning different neural networks to attitude decision and orbit decision, and sampling at different frequencies to ensure efficient acquisition of neural networks for attitude and orbit decisions. Attached Figure Description
[0030] Figure 1 is a training convergence curve of the attitude decision network according to an embodiment of the present invention;
[0031] Figure 2 is a training convergence curve of the orbit decision network according to an embodiment of the present invention;
[0032] Figure 3 is a diagram of the red-blue interaction process under network-driven implementation according to an embodiment of the present invention;
[0033] Figure 4 This is a flowchart of the algorithm in an embodiment of the present invention. Detailed Implementation
[0034] An intelligent decision-making method for collaborative optimization of spacecraft attitude and orbit is proposed. This method establishes a decision model for collaborative training of independent attitude and orbit decisions under a reinforcement learning mechanism and an asynchronous generation method for attitude and orbit training data, thereby enabling collaborative decision-making of attitude and orbit in time-sensitive and highly interactive scenarios.
[0035] Decision-making refers to generating expected values for spacecraft attitude and orbital changes, such as the expected attitude when the spacecraft orients itself to the sun, or the expected velocity increment at a specified position when changing its orbit. Control, on the other hand, refers to generating control commands for actuators such as momentum wheels and engines based on the expected values for attitude or orbital changes. This patent focuses on attitude and orbital decisions, specifically how to generate expected values for attitude and orbital changes. The specific steps are as follows:
[0036] (1) Based on the spacecraft orbital dynamics model and attitude dynamics model, the input of the orbital dynamics model is: the velocity increments in the x, y, and z directions under the orbital system, the spacecraft mass, the initial position and the initial velocity, and the output is: the position and velocity in the x, y, and z directions under the orbital system; the input of the attitude dynamics model is: the control torque, moment of inertia, initial position and initial velocity in the x, y, and z directions under the orbital system, and the output is: the four elements and Euler angular velocity in the x, y, and z directions under the orbital system.
[0037] (2) For the attitude and orbit joint optimization task, neural networks for attitude and orbit decisions are designed respectively, including the following design contents:
[0038] (2-1) Design an orbital decision neural network. The input of the network is the position and velocity of the spacecraft in the x, y, and z directions in the orbital system, the relative position of the sun in the orbital system, and other mission-related indicators, such as the effective observation time of time-sensitive events, the position of time-sensitive events in the orbital system, and the relative attitude and position of non-cooperative target operation parts. The output is the expected velocity increment in the x, y, and z directions in the orbital system.
[0039] (2-2) Design an attitude decision neural network. The inputs of the network are the spacecraft's position and velocity in the x, y, and z directions in the orbital frame, the relative position of the sun in the orbital frame, and other mission-related indicators, such as the effective observation time of time-sensitive events, the position of time-sensitive events in the orbital frame, the relative attitude and position of non-cooperative target operating parts, and the spacecraft's own attitude and attitude angular velocity. The output is the desired control torque in the x, y, and z directions in the orbital frame.
[0040] (2-3) Both the orbital decision neural network and the attitude decision neural network mentioned above are fully connected neural networks, and the activation function of the network is the ReLU function;
[0041] (2-4) Design reinforcement learning algorithms, including optimization of hyperparameters and design of reward functions. The reward function design comprises two parts: a public part and a private part. The public part evaluates the overall task completion rate, including cumulative illumination time and win rate, and these values are identical for training both neural networks. The private part represents the differences between the two neural networks during training. For the orbit control neural network, the private part is designed as a function related to velocity increment, relative position, and relative velocity, reflecting the influence of individual position on the overall task. For the attitude control neural network, the private part is designed as a function related to control torque, attitude quaternions, and attitude angular velocity, reflecting the influence of individual attitude on the overall task.
[0042] (3) Connect the data in the sample pool with the established reinforcement learning algorithm to realize the use of data such as relative position, relative velocity, absolute position, absolute velocity, attitude four elements, attitude angular velocity, fuel consumption, energy consumption, and solar azimuth angle generated by the spacecraft model for training of the neural network driven by reinforcement learning; and realize the use of the output of the neural network for iterative calculation of the spacecraft model.
[0043] Specifically, it includes the following steps:
[0044] (3-1) The orbit decision neural network generates the desired velocity increment, while the attitude decision neural network generates the desired torque. Due to the long-period cumulative effect of the orbit, the time period for generating the desired velocity increment is generally on the order of minutes, while the time period for generating the desired torque is on the order of seconds or even milliseconds. Therefore, when acquiring training data, an asynchronous data generation method for attitude and orbit is adopted: the iteration cycle of the spacecraft dynamics model is generally 64ms, while the cycle of attitude decision is generally on the order of seconds. The data used for training the attitude decision network is collected and trained at 20-100 times the iteration cycle of the spacecraft model, while the cycle of orbit decision is generally on the order of 10 minutes. The data used for training the orbit decision network is collected and trained at 10,000-30,000 times the iteration cycle of the spacecraft model. During the training process, the spacecraft's mass, moment of inertia, initial state, and other attribute parameters are randomly initialized to improve the generalization ability of the neural network.
[0045] (3-2) During training, the obtained neural network calculates attitude and orbit control variables, drives the spacecraft's orbital dynamics model and attitude dynamics model, and obtains the changes in the spacecraft's attitude and orbit over time from a certain initial state. Based on this data, the effectiveness of the attitude and orbit decisions generated by the trained neural network is evaluated. After repeated iterations, a successfully trained decision model is obtained.
[0046] (4) Collect the four elements of position, velocity, attitude, Euler angular velocity, mass, and moment of inertia of the spacecraft to be monitored for the predetermined mission target, and input them into the successfully trained spacecraft attitude-orbit coupling decision model to obtain the desired velocity increment and desired torque, and drive the spacecraft to change attitude and orbit.
[0047] The following is a further explanation with reference to specific embodiments:
[0048] In this embodiment, as shown in Figure 1, the intelligent decision-making method for coordinated optimization of spacecraft attitude and orbit includes the following specific steps:
[0049] (1) Using Python, a CW relative orbital dynamics model and a rigid body six-degree-of-freedom attitude dynamics model of the spacecraft were established. The inputs to the orbital dynamics model were: velocity increments in the x, y, and z directions under the orbital system, simulation time, and iteration step size. The outputs were: position and velocity in the x, y, and z directions under the orbital system. The inputs to the attitude dynamics model were: control torques in the x, y, and z directions under the orbital system, simulation time, and iteration step size. The outputs were: four elements and Euler angular velocities in the x, y, and z directions under the orbital system.
[0050] (2) Set the attribute parameters of the spacecraft. In this embodiment, two spacecraft are considered. Spacecraft 1 needs to approach within a certain range of spacecraft 2 in a specific attitude. Spacecraft 2 is the target of spacecraft 1.
[0051] Spacecraft 1:
[0052] Mass = 500 kg
[0053]
[0054] Initial position (0.01km): 1.3,0,0
[0055] Initial velocity (0.1 m / sec): 0, 0, 0
[0056] Spacecraft 2: Mass of 700 kg
[0057]
[0058] Initial position (0.01km): 0.7,0,0
[0059] Initial velocity (0.1 m / sec): 0, 0, 0
[0060] The initial position is randomly selected within a range of ±10km in the x, y, and z directions, and the initial velocity is randomly selected within a range of ±10m / s in the x, y, and z directions.
[0061] (3) For the attitude and orbit joint optimization task, neural networks for attitude and orbit decisions are designed respectively. For spacecraft 1, the attitude decision neural network and orbit decision neural network are designed as follows:
[0062] (3-1) The attitude decision neural network is a fully connected network containing two hidden layers, with the ReLU activation function. The network inputs are the position and velocity of spacecraft 1 in the x, y, and z directions in the CW coordinate system, the relative position of the sun in the CW coordinate system, the position and velocity of spacecraft 2 in the x, y, and z directions in the CW coordinate system, the attitude and attitude angular velocity of spacecraft 1, and the evaluation value of the impact of spacecraft 2 on spacecraft 1. The output is the expected velocity increment in the x, y, and z directions in the orbital system.
[0063] (3-2) The orbit decision neural network is a fully connected network containing two hidden layers. The activation function of the hidden layers is the ReLU function. The input of the network is the position and velocity of spacecraft 1 in the x, y, and z directions in the CW coordinate system, the relative position of the sun in the CW coordinate system, the position and velocity of spacecraft 2 in the x, y, and z directions in the CW system, and the evaluation value of the impact of spacecraft 1 on spacecraft 2.
[0064] (4) PPO (i.e., Proximity Policy Optimization) is adopted as the reinforcement learning algorithm. The parameters of PPO are designed as follows:
[0065] (4-1) The learning rate is chosen to be 2e -4 The clipping parameter is set to 0.1; the reward function is designed as follows:
[0066] For spacecraft 1, the reward function for orbit control is:
[0067]
[0068] Where cosθ is the angle between spacecraft 2 and the sun with spacecraft 1 as the vertex, and d 12 Let |Δv| be the distance between spacecraft 1 and spacecraft 2, |Δv| be the sum of the absolute values of the three-axis velocity increments in the orbital system, |ΔT| be the sum of the absolute values of the expected control torques of the spacecraft in the x, y, and z directions, and t+f be the common part of the reward function. t represents the planned mission execution time, which is equal to the current moment in a mission. By incorporating time into the reward function, the aim is to allow the spacecraft to complete the mission in the shortest possible time. f represents the planned mission completion rate, which is equal to the continuous effective observation time of the planned mission target. Effective observation of the target is only possible when the relative position and attitude simultaneously meet certain constraints.
[0069] For spacecraft 1, the reward function for attitude control is:
[0070] rew0b2 +=C2+|ΔT|+t+f
[0071] Where C2 is the impact assessment value of spacecraft 2 on spacecraft 1, |ΔT| is the sum of the absolute values of the expected control torques of the spacecraft in the x, y, and z directions, and t+f is the common part of the reward function. t represents the mission execution time, which is equal to the current moment in a mission. By incorporating time into the reward function, the aim is to allow the spacecraft to complete the mission in the shortest possible time. f represents the mission completion rate, which is equal to the continuous effective observation time of the predetermined mission target. Effective observation of the target is only possible when the relative position and attitude simultaneously meet certain constraints.
[0072] (5) Connect the spacecraft model established in steps 1 and 2 with the reinforcement learning algorithm established in steps 3 and 4.
[0073] (6) Run the engineering code generated in step (5) to train the neural network. During the training process, the spacecraft's mass, moment of inertia, initial state and other attribute parameters are randomly initialized to improve the generalization ability of the neural network.
[0074] (7) After training, several sets of spacecraft attribute parameters are randomly generated. The attitude and orbit control quantities are calculated using the neural network obtained in step 6 to drive the spacecraft model, thereby obtaining the attitude and orbit changes of the spacecraft over time from a certain initial state, and evaluating the effectiveness of the attitude and orbit decisions generated by the trained neural network.
[0075] Figure 1 shows the convergence curve of the attitude decision network training, and Figure 2 shows the convergence curve of the orbit decision network training. The vertical axis represents the cumulative value of the spacecraft reward function. As the number of training rounds increases, the cumulative reward of the spacecraft gradually increases, converging to the maximum value after 10,000 rounds. Figure 3 shows the interaction process between spacecraft 1 and spacecraft 2. Figure 3(a) shows the trajectory of the spacecraft in the CW coordinate system, Figure 3(b) shows the relative distance between spacecraft 1 and spacecraft 2, and Figure 3(c) shows the angle between the line connecting spacecraft 1 and spacecraft 2 and the direction of the sun. It can be seen that spacecraft 1 eventually arrives within 50km of spacecraft 2 in a specific direction for escort flight. The model obtained by training using the method proposed in this invention can realize the collaborative decision-making of spacecraft attitude and orbit in time-sensitive scenarios.
[0076] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.
Claims
1. An intelligent decision-making method for spacecraft attitude and orbit coordination, characterized in that, include: Conduct spacecraft orbital dynamics and attitude dynamics simulations, and obtain the spacecraft's position, velocity, attitude, and Euler angular velocity relative to the predetermined mission target with a predetermined simulation iteration step size; The data obtained above and the relative position of the Sun in the orbital system are used as the sample pool; A spacecraft attitude and orbit coupling decision model is established, including an attitude decision neural network, an orbit decision neural network, and a reward function algorithm for reinforcement learning of the above two decision neural networks. The attitude decision neural network and the orbit decision neural network generate the expected control torque and velocity increment of the spacecraft in the x, y, and z directions under the orbital system, respectively, to control the attitude and orbit of the spacecraft. Sample pool data is asynchronously collected at different sampling frequencies and used as training data for attitude decision neural network and orbit decision neural network, respectively. The weights of the two decision neural networks are iteratively updated to complete the network training and obtain a successfully trained spacecraft attitude-orbit coupling decision model. The system collects the position, velocity, attitude, Euler angular velocity, and relative position of the spacecraft to be monitored relative to the target, as well as the Sun's position in the orbital system. This data is then fed into a successfully trained spacecraft attitude-orbit coupling decision model to obtain the desired velocity increment and desired torque. This data drives the spacecraft to change its attitude and orbit, thereby achieving an orbital approach to the target. The reward function algorithm consists of a public part and a private part: The public part is used to evaluate the completion of the predetermined task, and the same value is used for training the attitude decision neural network and the trajectory decision neural network; The private part differs in its training for the attitude decision neural network and the orbit decision neural network: for the orbit decision neural network, the private part is designed as a function related to the velocity increment, while for the attitude decision neural network, the private part is designed as a function related to the control torque.
2. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, Establishing a neural network for trajectory decision-making includes: An orbital decision neural network is established. The input data of the network includes the position and velocity of the spacecraft relative to the predetermined mission target and the relative position of the sun in the orbital system. The output is the expected velocity increment of the spacecraft in the x, y, and z directions in the orbital system.
3. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, Establish a neural network for attitude decision-making, including: Design an attitude decision neural network. The input data of the network includes the attitude quaternion, attitude angular velocity, position, and velocity of the spacecraft relative to the predetermined mission target. The output is the expected control torque of the spacecraft in the x, y, and z directions in the orbital system.
4. The intelligent decision-making method for spacecraft attitude and orbit coordination according to any one of claims 1-3, characterized in that, Both the orbital decision neural network and the attitude decision neural network are fully connected neural networks, and the activation function of both networks is the ReLU function.
5. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, The reward function algorithm for the orbital decision neural network is as follows: In the formula, Let d be the angle between the spacecraft as the apex, the predetermined mission target, and the sun. 12 The distance between the spacecraft and its intended mission target. It is the sum of the absolute values of the velocity increments in the x, y, and z directions within the orbital system. This is the sum of the absolute values of the desired control torques of the spacecraft in the x, y, and z directions. Indicates the scheduled time for the task. Indicates the completion rate of the scheduled task. This is the public part of the reward function.
6. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, The reward function algorithm for the posture decision neural network is as follows: in The impact assessment value of the predetermined mission objectives on the spacecraft. This is the sum of the absolute values of the desired control torques of the spacecraft in the x, y, and z directions. Indicates the time elapsed since the task began. Indicates the task completion rate. This is the public part of the reward function.
7. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, Training data for the attitude decision neural network was collected from the sample pool at a time interval of 20-100 times the simulation iteration step size, and training data for the orbit decision neural network was collected from the sample pool at a time interval of 10,000-30,000 times the simulation iteration step size.
8. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, During network training, the spacecraft's mass, moment of inertia, and initial state are randomly initialized to improve the generalization ability of the attitude decision neural network and the orbit decision neural network.
9. The intelligent decision-making method for spacecraft attitude and orbit coordination according to claim 1, characterized in that, The spacecraft's moment of inertia, initial position, initial velocity, and velocity increments in the x, y, and z directions within the orbital frame are input into the spacecraft's orbital dynamics model. With a predetermined simulation iteration step size, the spacecraft's position and velocity relative to the predetermined mission target are obtained. The spacecraft's mass, initial position, initial velocity, and control torques in the x, y, and z directions within the orbital frame are input into the spacecraft's attitude dynamics model. With a predetermined simulation iteration step size, the spacecraft's attitude four elements and Euler angular velocity relative to the predetermined mission target are obtained.