Spacecraft attitude stability control method based on liquid neural network

Abstract

The invention discloses a spacecraft attitude stability control method based on a liquid neural network. The method comprises the following steps: S1, establishing a spacecraft attitude dynamic model under multi-source complex disturbance; s2, designing a conventional spacecraft attitude stability controller, and constructing a data set in combination with the attitude dynamics model; s3, training and obtaining a liquid neural network interference observer by adopting a liquid time constant network and a neural circuit strategy architecture; s4, designing a spacecraft attitude stability controller based on a liquid neural network disturbance observer, and proving the stability of a closed-loop system by using a Lyapunov function; s5, verifying the estimation effect of the liquid neural network interference observer and the stability effect of the spacecraft attitude control system through simulation; results show that the method has the advantages that interference of the spacecraft attitude can be effectively estimated, stable control over the spacecraft attitude can be achieved, high real-time performance, low power consumption and high robustness are achieved, and the spacecraft attitude stable control efficiency can be improved.

Description

Technical Field

[0001] This invention belongs to the field of spacecraft control technology, and relates to a spacecraft attitude control method, particularly a spacecraft attitude stabilization control method based on a liquid neural network. Background Technology

[0002] With the continuous expansion of space exploration and the rapid development of aerospace technology, spacecraft on-orbit missions are becoming increasingly complex, placing ever higher demands on the performance of attitude control systems. However, spacecraft attitude control systems face numerous challenges, such as unknown model parameters, difficulty in measuring state information, actuator failures, and computational resource limitations. These factors severely impact the performance of traditional control methods and may even lead to mission failure.

[0003] To address the aforementioned problems, many modern attitude control methods have been proposed, including adaptive control, sliding mode control, and H∞ control. Among these, the control method based on disturbance observers is one of the most promising approaches for spacecraft attitude control. However, traditional disturbance observers cannot accurately estimate external disturbances, and these methods have relatively slow response times.

[0004] While deep neural networks, such as RNNs and Transformers, have shown advantages in certain scenarios, they still have many drawbacks. For example, the high inference latency of GPU deployments makes it difficult to meet the real-time requirements of control loops; once the weights are trained, they cannot be modified, making it difficult to adaptively adjust the internal state and adapt to dynamic changes online, requiring retraining on the ground if mission requirements change; the hardware devices that traditional neural networks rely on consume a lot of power, far exceeding the power supply capacity of spacecraft; they lack memory and have limited adaptability to dynamic systems; traditional neural networks are like black boxes, and their lack of interpretability is one of the important reasons limiting their application in the spacecraft field.

[0005] The liquid dynamic neural network based on a sparse connection structure, by introducing continuous-time dynamics and recursive connection structures, possesses the ability to process temporal information and dynamic memory, providing a new approach for the intelligent control of complex dynamic systems. As a novel neural network structure, its variable weights enable it to excel in processing time-varying and dynamic data, adapting to constantly changing environments and accomplishing complex tasks with fewer neurons. It not only boasts low computational energy consumption but also significantly improves the interpretability of the neural network due to its simple structure. The liquid neural network adjusts the neuron state vectors and connection weights in real time by accepting new disturbance inputs. Employing nonlinear activation functions, it utilizes the model's memory mechanism, dynamic adjustment mechanism, and recursive feedback mechanism to analyze spacecraft attitude correction strategies. Its high efficiency and low energy consumption give it unparalleled advantages in the field of spacecraft control. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a highly adaptive, low-power spacecraft attitude stabilization control method based on liquid neural networks. This method enables adaptive attitude stabilization control of spacecraft under conditions of external disturbances, significantly reducing the modeling accuracy requirements of traditional control systems and the high energy consumption of traditional neural networks. It meets the needs of spacecraft for high-precision, dynamic adjustment, and adaptive control, thereby improving spacecraft attitude control efficiency and extending the on-orbit lifespan of satellites. This invention is achieved using the following technical solution: S1. Under the complex disturbances of multiple sources, such as model parameter uncertainty, external interference, and actuator failure, establish a spacecraft attitude dynamics model; S2. Design a conventional spacecraft attitude stabilization controller under the above-mentioned multi-source complex disturbances, combine the attitude dynamics model to obtain error, control and disturbance information, and integrate the above information to construct a dataset; S3. Using a liquid time constant network and neural circuit strategy architecture, a liquid neural network disturbance observation model is trained using the above dataset, and a liquid neural network disturbance observer is constructed. S4. Design a spacecraft attitude stabilization controller based on a liquid neural network disturbance observer, and prove the stability of the closed-loop system using Lyapunov functions. S5. The estimation effect of this liquid neural network disturbance observer and the stabilization effect of the spacecraft attitude control system are verified through simulation. The specific process of establishing the spacecraft attitude dynamics model in step S1 is as follows: under the complex disturbances of multiple sources such as model parameter uncertainty, external interference, and actuator failure, the spacecraft attitude angle, attitude angular velocity, external disturbance torque, inertial parameters, control torque and unmodeled error upper limit are considered to establish the spacecraft attitude dynamics model. The specific process of obtaining the dataset under the conventional spacecraft attitude control system in step S2 is as follows: under the multi-source complex disturbances of model parameter uncertainty, external interference, and actuator failure, a conventional spacecraft attitude stabilization controller is first designed. The controller is used to realize the spacecraft attitude stabilization control. Combined with the attitude dynamics model, the control history is recorded to obtain attitude angle error, attitude angular velocity error, control torque and random disturbance information. The above information is integrated to construct the dataset. The construction of the liquid neural network interference observer in step S3 is specifically as follows: A liquid time constant network and neural circuit strategy architecture are adopted. The attitude angle error, attitude angular velocity error, and control torque in the dataset are used as inputs, and the interference estimate is used as the output. A liquid neural network interference observation model is trained, and a liquid neural network interference observer is constructed based on this model. The interference observer is characterized by the continuous-time dynamics of the liquid time constant network and the sparse connections of the neural circuit strategy. The neural circuit strategy is a wiring pattern inspired by biological neurons, integrating multiple liquid time constant network neurons into a multi-layer structure to achieve sparse connections between sensory neurons, inter-neuronal layers, command neurons, and motor neurons. The liquid time constant network is a dynamic neural network constructed based on differential equations. Its core lies in using dynamic adjustment of the time constant to simulate the adaptability of biological synapses. This characteristic is based on the STDP rule, a biomimetic learning rule that simulates the synaptic strength in the biological brain. It has a dynamic adjustment mechanism, capable of updating the synaptic state over time and adjusting the connection weights between neurons in real time to adapt to changes in external input. That is, it dynamically adjusts the interference estimate in real time according to the spacecraft attitude deviation. The mathematical expression of the weights is: In the formula, Synaptic weights vary depending on the temporal sequence and time difference of the impulses from the preceding and following neurons. The time difference between the preceding and following pulses. The learning rate parameter controls the magnitude of the enhancement reduction. The time window constant determines the time range of the influence; The liquid neural network model possesses memory properties, based on dynamic neurons derived from differential equations. By simulating the continuous-time dynamic behavior of real neurons and combining the cumulative effect of historical data over time with current input data, the prediction results are dynamically adjusted. The liquid neural network adjusts the neuron state vector and connection weights in real time by accepting new disturbance inputs. Employing a nonlinear activation function, it utilizes the model's memory mechanism, dynamic adjustment mechanism, and recursive feedback mechanism to estimate the disturbances corresponding to spacecraft attitude errors. The state equation of the liquid neuron is: In the formula, It is the neuron state vector. It is a time constant. It is an input weight matrix that is adjusted online using the STDP rule. It is the input signal. It is a recursive weight matrix. It is a bias term. It is a non-linear activation function; The design of the spacecraft attitude stabilization controller in step S4 is based on a liquid neural network disturbance observer, and the stability of the closed-loop system is proved by Lyapunov function. The specific process is as follows: the disturbance estimate based on the liquid neural network disturbance observer is input to the front end of the controller to introduce equivalent compensation, reduce the influence of the comprehensive disturbance, and the system stability is proved by Lyapunov function. The process of verifying the estimation effect of the liquid neural network interference observer and the stabilization effect of the attitude stabilization system through simulation in step S5 is as follows: random interference is input into the spacecraft attitude stabilization system, and the liquid neural network interference observer obtains the interference estimate based on the attitude angle error, attitude angular velocity error and historical control torque input of the spacecraft. Based on this estimate, equivalent compensation is introduced in the controller to verify that the interference estimate of the observer is small compared with the actual interference error and that the system always maintains a stable state.

[0007] The core of liquid neural networks is to process time-series data using a continuous-time dynamic system within a recursive network. Mathematically, this is expressed as using differential equations to mimic the changes in the state of biological neurons over time. The steps for building the liquid neural network in step S3 are as follows: Step S3.1: Network Architecture Design and Initialization. Before building the liquid neural network, the network structure is first determined, including the input layer, the liquid layer, and the output layer. The liquid layer consists of dynamic neurons, with connection weights and dynamic parameters randomly initialized to promote training diversity and stability. The liquid layer contains liquid neural network neurons, including sensory neurons, interneuronal layers, command neurons, and motor neurons.

[0008] Step S3.2: Data Preprocessing and Input Sampling. Input time-series data samples are selected from the training set and normalized. Since the liquid neural network relies on temporal dynamics, the input data needs to be expanded step-by-step to preserve time dependencies.

[0009] Step S3.3: Forward Propagation and Dynamic State Update. As input data passes through the liquid layer, the dynamic state of the neuron is controlled by a differential equation, whose state equation is: In the formula, It is the neuron state vector. It is a time constant. It is an input weight matrix that is adjusted online using the STDP rule. It is the input signal. It is a recursive weight matrix. It is a bias term. It is a non-linear activation function; Step S3.4: Nonlinear Activation and Output Generation. The output of the liquid layer is transformed and passed to the output layer through a nonlinear activation function. The output layer designs and uses functions according to task requirements to generate the final prediction result.

[0010] Step S3.5: Loss Function Calculation. Spacecraft attitude stabilization control is a time-series task, therefore, an additional time-dimensional loss is introduced by accumulating the error along the time step. The error between the predicted value and the true value is compared to optimize the network model in reverse.

[0011] Step S3.6: Time-dependent backpropagation. Calculate the gradient of the loss with respect to the parameters using the time-unfolded network. Due to the dynamic nature of LNNs, differentiation along the time chain is required, and the weights and time constants are adjusted using gradient descent. Dynamic parameters, etc.

[0012] Step S3.7: Parameter Optimization and Update. Update parameters using an optimization algorithm. Update physical constraints (such as...) in dynamic parameter updates through projection. (This is a consideration.)

[0013] Step S2.8: Iterative Training and Model Evaluation. Repeat steps S3.2 to S3.7 until the loss converges. Evaluate the model performance on the validation set and adjust the hyperparameters. Finally, test the liquid neural network model on the test set.

[0014] The beneficial effects of this invention are: a liquid time constant network-neural circuit strategy architecture model is established, realizing the estimation and compensation of disturbances. Through this method, stable attitude control of spacecraft can be achieved under conditions of external disturbances, uncertain model parameters, and actuator failures. This significantly reduces the modeling accuracy requirements of traditional control and the high energy consumption and low interpretability of traditional neural networks, meeting the needs of high-precision, dynamic adjustment, and adaptive control of spacecraft, and helping to improve the efficiency of spacecraft attitude control. Attached Figure Description

[0015] Figure 1 The flowchart of the spacecraft attitude stabilization control method based on liquid neural network described in this invention patent; Figure 2 This is the flowchart of the liquid neural network construction process described in this invention patent; Figure 3 This is a flowchart of the algorithm for attitude control based on liquid neural networks as described in this invention patent; Figure 4 This is a structural diagram of the spacecraft closed-loop attitude control system described in this invention patent. Detailed Implementation The present invention will now be described in detail with reference to the accompanying drawings. Establishment as follows: Figure 1 A spacecraft attitude stabilization control method based on liquid neural networks.

[0016] Step S1: Under the complex disturbances of multiple sources, such as model parameter uncertainty, external interference, and actuator failure, establish a spacecraft attitude dynamics model, including spacecraft attitude angle, attitude angular velocity, external disturbance torque, upper limit of unmodeled error, inertial parameters, and control torque information. Step S2: Based on the principles of spacecraft attitude and orbit control, select the deviations between the parameters of the spacecraft after attitude changes caused by the involuntary control moment of motion during its on-orbit mission and the ideal attitude parameters. These parameters include physical quantities such as attitude angle, attitude angular velocity, and vibration modes, which can be measured by the onboard inertial measurement unit. Since the spacecraft controller cannot directly obtain the disturbance values ​​that cause attitude deviations, it is necessary to estimate the disturbance values ​​based on the attitude deviations. Under the complex disturbances of multiple sources, including model parameter uncertainty, external interference, and actuator failure, a conventional spacecraft attitude stabilization controller is designed first. This controller is used to achieve spacecraft attitude stabilization control. Combined with the attitude dynamics model, the state error information during the control process under disturbances—attitude angle error, attitude angular velocity error, control torque, and external disturbance information—is obtained. The above data is used to form a training dataset, providing a basis for model training of the liquid neural network.

[0017] Step S3: Using the constructed dataset, adopt a liquid time constant network and neural circuit strategy architecture, take the attitude angle error, attitude angular velocity error and control torque in the dataset as input, and the disturbance estimate as output. 75% of the data is used as the training set, 15% as the validation set and 10% as the test set to train the liquid neural network disturbance observation model, and build a liquid neural network disturbance observer based on this model. After training, a usable liquid neural network model that converts state error into equivalent disturbance estimate is obtained. The liquid time constant network combined with the neural circuit strategy architecture is characterized by sparse connectivity and continuous-time dynamics. The neural circuit strategy is a wiring pattern inspired by biological neurons, integrating multiple liquid time constant network neurons into a multi-layer structure to achieve sparse connections between sensory neurons, interneuronal layers, command neurons, and motor neurons. The liquid time constant network is a dynamic neural network constructed based on differential equations. Its core lies in using dynamic adjustment of the time constant to simulate the adaptability of biological synapses. This characteristic is based on the STDP rule, a biomimetic learning rule that simulates synaptic strength in the biological brain. It has a dynamic adjustment mechanism that can update the synaptic state over time and adjust the connection weights between neurons in real time to adapt to changes in external input. That is, it dynamically adjusts the estimated value of the disturbance in real time according to the spacecraft attitude deviation. The mathematical expression of the weights is: In the formula, Synaptic weights vary depending on the temporal sequence and time difference of the impulses from the preceding and following neurons. The time difference between the preceding and following pulses. The learning rate parameter controls the magnitude of the enhancement reduction. The time window constant determines the time range of the influence; The liquid neural network model possesses memory properties, based on dynamic neurons derived from differential equations. By simulating the continuous-time dynamic behavior of real neurons and combining the cumulative effect of historical data over time with current input data, the prediction results are dynamically adjusted. The liquid neural network adjusts the neuron state vector and connection weights in real time by accepting new disturbance inputs. Employing a nonlinear activation function, it utilizes the model's memory mechanism, dynamic adjustment mechanism, and recursive feedback mechanism to estimate the disturbances corresponding to spacecraft attitude errors. The state equation of the liquid neuron is: In the formula, It is the neuron state vector. It is a time constant. It is an input weight matrix that is adjusted online using the STDP rule. It is the input signal. It is a recursive weight matrix. It is a bias term. It is a non-linear activation function; Step S4: Design a spacecraft attitude stabilization controller based on a liquid neural network disturbance observer, and prove the stability of the closed-loop system using the Lyapunov function. The specific process is as follows: input the disturbance estimate based on the liquid neural network disturbance observer into the front end of the controller to introduce equivalent compensation, reduce the influence of the comprehensive disturbance, and prove the system stability using the Lyapunov function. Step S5: Estimate the disturbance using the spacecraft attitude stabilization controller based on the liquid neural network disturbance observer and perform simulation verification. The specific process is as follows: input random disturbance into the attitude system, and the liquid neural network disturbance observer obtains the disturbance estimate based on the spacecraft's attitude angle error, attitude angular velocity error and historical control torque input. Based on this estimate, an equivalent compensation is introduced in the controller. The disturbance estimate of the observer is verified to have a small error with the actual disturbance, and the system always maintains a stable state.

[0018] It should be noted that the method of this invention has a wide range of applications, not only applicable to the attitude control of traditional spacecraft, but also to the stability control of future ultra-large space infrastructure in-orbit construction. Specific embodiments have been used to illustrate the principles and implementation of this invention. The descriptions of these embodiments are merely to help understand the method and its core ideas; furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention. Those skilled in the art will realize that the embodiments described herein are to help the reader understand the principles of this invention and should be understood as not limiting the scope of protection of this invention to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the essence of this invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A spacecraft attitude stabilization control method based on liquid neural networks, characterized in that, Includes the following steps: S1. Under the complex disturbances of multiple sources, such as model parameter uncertainty, external interference, and actuator failure, establish a spacecraft attitude dynamics model; S2. Design a conventional spacecraft attitude stabilization controller under the above-mentioned multi-source complex disturbances, combine the attitude dynamics model to obtain error, control and disturbance information, and integrate the above information to construct a dataset; S3. Using a liquid time constant network and neural circuit strategy architecture, a liquid neural network disturbance observation model is trained using the above dataset to construct a liquid neural network disturbance observer. S4. Design a spacecraft attitude stabilization controller based on a liquid neural network disturbance observer, and prove the stability of the closed-loop system using Lyapunov functions. S5. The estimation effect of this liquid neural network disturbance observer and the stabilization effect of the spacecraft attitude control system are verified through simulation. The specific process of establishing the spacecraft attitude dynamics model in step S1 is as follows: under the complex disturbances of multiple sources such as model parameter uncertainty, external interference, and actuator failure, the spacecraft attitude angle, attitude angular velocity, external disturbance torque, inertial parameters, control torque and unmodeled error upper limit are considered to establish the spacecraft attitude dynamics model. The specific process of obtaining the dataset under the conventional spacecraft attitude control system in step S2 is as follows: under the multi-source complex disturbances of model parameter uncertainty, external interference, and actuator failure, a conventional spacecraft attitude stabilization controller is first designed. The controller is used to realize the spacecraft attitude stabilization control. Combined with the attitude dynamics model, the control history is recorded to obtain attitude angle error, attitude angular velocity error, control torque and random disturbance information. The above information is integrated to construct the dataset. The specific process of constructing the liquid neural network disturbance observer in step S3 is as follows: Using a liquid time constant network and neural circuit strategy architecture, the attitude angle error, attitude angular velocity error, and control torque from the dataset are used as inputs, and the disturbance estimate is used as the output. A liquid neural network disturbance observation model is trained, and a liquid neural network disturbance observer is constructed based on this model; disturbance... The observer is characterized by the continuous-time dynamics of the liquid time constant network and the sparse connections of the neural circuit strategy. The neural circuit strategy is a wiring pattern inspired by biological neurons, integrating multiple liquid time constant network neurons into a multi-layered structure to achieve sparse connections between sensory neurons, interneuronal layers, command neurons, and motor neurons. The liquid time constant network is a dynamic neural network constructed based on differential equations. Its core lies in simulating the adaptability of biological synapses by dynamically adjusting the time constant. This characteristic is based on the STDP rule, a biomimetic learning rule that simulates synaptic strength in the biological brain. It has a dynamic adjustment mechanism, capable of updating synaptic states over time and adjusting the connection weights between neurons in real time to adapt to changes in external input. Specifically, it dynamically adjusts the estimated value of interference based on the spacecraft's attitude deviation in real time. The mathematical expression of the weights is as follows: In the formula, Synaptic weights vary depending on the temporal sequence and time difference of the impulses from the preceding and following neurons. The time difference between the preceding and following pulses. The learning rate parameter controls the magnitude of the enhancement reduction. The time window constant determines the time range of the influence; The liquid neural network model possesses memory properties, based on dynamic neurons derived from differential equations. By simulating the continuous-time dynamic behavior of real neurons and combining the cumulative effect of historical data over time with current input data, the prediction results are dynamically adjusted. The liquid neural network adjusts the neuron state vector and connection weights in real time by accepting new disturbance inputs. Employing a nonlinear activation function, it utilizes the model's memory mechanism, dynamic adjustment mechanism, and recursive feedback mechanism to estimate the disturbances corresponding to spacecraft attitude errors. The state equation of the liquid neuron is: In the formula, It is the neuron state vector. It is a time constant. It is an input weight matrix that is adjusted online using the STDP rule. It is the input signal. It is a recursive weight matrix. It is a bias term. It is a non-linear activation function; The design of the spacecraft attitude stabilization controller in step S4 is based on a liquid neural network disturbance observer, and the stability of the closed-loop system is proved by Lyapunov function. The specific process is as follows: the disturbance estimate based on the liquid neural network disturbance observer is input to the front end of the controller to introduce equivalent compensation, reduce the influence of the comprehensive disturbance, and the system stability is proved by Lyapunov function. The process of verifying the estimation effect of the liquid neural network interference observer and the stabilization effect of the attitude stabilization system through simulation in step S5 is as follows: random interference is input into the spacecraft attitude stabilization system, and the liquid neural network interference observer obtains the interference estimate based on the attitude angle error, attitude angular velocity error and historical control torque input of the spacecraft. Based on this estimate, equivalent compensation is introduced in the controller to verify that the interference estimate of the observer is small compared with the actual interference error and that the system always maintains a stable state.

Images