Aircraft RCS intelligent fault-tolerant distribution method based on neural network

Abstract

The invention discloses an aircraft RCS intelligent fault-tolerant distribution method based on a neural network, and belongs to the technical field of aerospace. The implementation method comprises the following steps: establishing a three-degree-of-freedom attitude kinetic equation of the aircraft; based on the aircraft three-degree-of-freedom attitude kinetic equation, establishing an aircraft RCS control distribution model; the method comprises the following steps: converting an aircraft RCS control distribution problem into an optimization problem for solving, fusing a high-practicability tracking task instruction and a random instruction covering an RCS full operation space, taking a nozzle health state as one of inputs, systematically introducing a fault mode, and designing and generating a data set for training a neural network; and constructing a neural network, and training the neural network by using the generated data set to obtain the RCS intelligent fault-tolerant distributor. The RCS intelligent fault-tolerant distributor is applied to an actual aircraft attitude control task, a switching time instruction of each spray pipe of the RCS is generated, it is ensured that RCS output can meet the control instruction requirement, and fault-tolerant distribution control over the aircraft attitude is achieved.

Description

Technical Field

[0001] This invention belongs to the field of aerospace technology and relates to a neural network-based intelligent fault-tolerant allocation method for RCS of aircraft. Background Technology

[0002] When flying in space or the upper atmosphere, aircraft typically utilize a Reaction Control System (RCS) to achieve effective attitude control. The RCS generates reaction torque through multiple propulsion nozzles distributed at different locations on the aircraft. Therefore, it is necessary to rationally allocate the opening and closing times of each nozzle so that the RCS output is as close as possible to the attitude control system's commands. Furthermore, during high-dynamic flight, the RCS may experience complex failures due to various reasons such as equipment aging and pipe blockage, including varying degrees of efficiency degradation, complete failure of a single nozzle, or even multiple nozzle failures. Therefore, designing a fault-tolerant allocation method that can adapt to different failure conditions and maximize the aircraft's control capability to complete attitude control tasks is crucial to the success or failure of the flight. Currently, there are several typical methods for the RCS control allocation problem: the table lookup method is fast, but the allocation method is too fixed, has weak adaptability, and is difficult to cope with the complex and ever-changing actual flight environment; the pseudo-inverse method has low computational cost and is easy to deploy and apply online, but it is difficult to meet the specific constraints of the control variables, such as the non-negativity of the switching time; the optimization-based allocation algorithm transforms the allocation problem into a constrained optimization problem, which can explicitly handle various constraints and pursue optimal performance to a certain extent, but it has high computational complexity, long time consumption, and is difficult to call online at high frequency. Summary of the Invention

[0003] The main objective of this invention is to provide a neural network-based intelligent fault-tolerant allocation method for aircraft RCS (Rapid Cross Section). By establishing an aircraft attitude dynamics model and an RCS control allocation model, data samples are obtained by solving the allocation problem using optimization methods. Highly practical tracking mission commands and random commands covering the entire RCS operating space are integrated, and the nozzle health status is used as one of the inputs. Fault modes are systematically introduced, enabling the generated dataset to effectively train the neural network to adaptively learn the optimal allocation strategy under various fault modes. This results in an intelligent RCS allocator with high precision, strong generalization, and fault tolerance, enhancing the fault-tolerant allocation capability for aircraft attitude control.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] This invention discloses a neural network-based intelligent fault-tolerant allocation method for aircraft RCS (Rapid Cross Section) control. It establishes an aircraft attitude dynamics model and an RCS control allocation model, obtains data samples by solving the allocation problem using optimization methods, and generates impulse moment commands using a hybrid strategy. This strategy integrates highly practical tracking mission commands and random commands covering the entire RCS operating space, and systematically considers nozzle damage conditions to ensure the neural network can learn allocation strategies under different fault conditions. Furthermore, a neural network is designed and trained to obtain an intelligent fault-tolerant RCS allocator. The intelligent fault-tolerant RCS allocator is applied to actual aircraft attitude control tasks to generate on / off time commands for each nozzle in the RCS, ensuring that the RCS output meets the control command requirements and achieving fault-tolerant allocation control of the aircraft attitude.

[0006] This invention discloses a neural network-based intelligent fault-tolerant allocation method for aircraft RCS, comprising the following steps:

[0007] Step 1: Establish the three-degree-of-freedom attitude dynamics equations for the aircraft.

[0008] The three-degree-of-freedom attitude dynamics equations of the aircraft during flight are expressed as follows:

[0009]

[0010] in, Let γ be the Euler angle vector of the aircraft. ψ represents the roll angle, pitch angle, and yaw angle of the aircraft, respectively; ω = [ω x ,ω y ,ω z ] T Let ω be the three-axis attitude angular velocity vector of the aircraft. × I is the skew-symmetric matrix corresponding to ω; I is the aircraft's moment of inertia matrix; R is the transformation matrix, specifically expressed as follows:

[0011]

[0012] M is the torque acting on the aircraft, specifically expressed as follows:

[0013] M = M base +M ctrl +f (3)

[0014] Among them, M base The fundamental torques of an aircraft include the fundamental aerodynamic torque of the fuselage, the gravity gradient torque, etc.; M ctrl The control torque generated by the aircraft's control actuators; f represents various disturbances and uncertainties. For an aircraft that uses only RCS for attitude control, the thrust and torque generated when each nozzle of the RCS is open are fixed, and the aircraft's control torque is the sum of the torques output by each nozzle:

[0015]

[0016] Where n is the number of nozzles; M ctrl,i M represents the torque output by the i-th nozzle (i = 1, 2, ..., n); ri Let be the torque generated when the i-th nozzle is opened.

[0017] Step 2: Based on the three-degree-of-freedom attitude dynamics equations of the aircraft, establish the RCS control allocation model of the aircraft.

[0018] For an RCS system with n nozzles, the control efficiency matrix of the RCS system is defined as follows:

[0019] B rcs =[M r1 M r2 ,...,M ri ,...,M rn ], i=1,2,...,n (5)

[0020] Considering the discrete torque output characteristics of RCS, the allocation essentially involves calculating the opening time of each nozzle within a specified time period Δt, so that the sum of the impulse torques generated by each nozzle is as close as possible to the control command torque M. cmd The corresponding impulse torque command H cmd =M cmd Δt, that is:

[0021] B rcs t rcs →H cmd (6)

[0022] Among them, t rcs =[t r1 ,t r2 ,...,t rn ] T This indicates the opening time of each nozzle.

[0023] Step 3: Transform the aircraft RCS control allocation problem into an optimization problem, solve the RCS control allocation problem, and design and generate a dataset for training the neural network.

[0024] Considering the non-negativity of nozzle opening time, the RCS control allocation problem is transformed into the following optimization problem:

[0025] Therefore, given B rcs and H cmd At that time, the non-negative least squares method is used to solve for the opening time of each nozzle, and a dataset sample is generated.

[0026] When generating the dataset for training the RCS intelligent fault-tolerant assignment neural network, a hybrid strategy is employed to generate the impulse moment instruction set {H}. cmd This strategy combines two methods: generating instructions that track the nominal trajectory and generating random instructions that cover the entire RCS operating space.

[0027] {H cmd}={H cmd , track}∪{H cmd , rand} (8)

[0028] The command to track the nominal trajectory originates from the impulse moment command H generated by the aircraft tracking the reference attitude profile during the actual mission. cmd,track Specifically, it is described as follows:

[0029] H cmd , track =f(Ω) ref ,ω ref ,ΔΩ init ,d) (9)

[0030] Among them, Ω ref and ω ref These are the reference Euler angles and angular velocity, ΔΩ, respectively. init Let d represent the initial attitude deviation, and d represent external disturbances and system uncertainties.

[0031] Random instruction generation covering the entire RCS operating space refers to instruction generation based on the control efficiency matrix B. rcs Determine the boundary of the impulse moment that the RCS can generate within a specified time period Δt [H] min H max Within this boundary, impulse moment command H is generated by uniformly and randomly sampling. cmd,rand :

[0032] H cmd,rand ~U(H min H max (10) Among them,

[0033]

[0034] The input to the dataset samples for designing the RCS intelligent fault-tolerant assignment neural network includes a 3D impulse moment command H. cmd and n-dimensional nozzle health status:

[0035] g rcs =[g r1 ,g r2 ,...,g ri ,...,g rn ] T , gri ∈[0,1], i=1,2,...,n (12)

[0036] Among them, g ri A value of 1 indicates that the i-th nozzle is working normally; g ri A value of 0 indicates a complete failure of the i-th nozzle, resulting in zero output torque, corresponding to matrix B. rcs Set the i-th column of g to zero; ri The values ​​between 0 and 1 represent the degree of decrease in nozzle efficiency and output torque. The output of the dataset samples is the solved nozzle opening time t. rcs .

[0037] When generating the dataset, nozzle damage is considered, i.e., g is modified. rcs The values ​​of the elements in the middle characterize single nozzle failure, multi-nozzle failure, and different degrees of nozzle efficiency degradation, ensuring that the neural network can learn allocation strategies under different failure conditions and enhance fault-tolerant allocation capabilities.

[0038] Step 4: Construct a neural network. Train the neural network using the dataset obtained in Step 3 to obtain the RCS intelligent fault-tolerant allocator, which enhances the fault-tolerant allocation capability for aircraft attitude control.

[0039] Design the structure of the neural network, with the input layer including a 3-dimensional impulse moment instruction H. cmd and n-dimensional nozzle health status g rcs The output layer includes the opening time t of each nozzle. rcs Optionally, a fully connected neural network structure is adopted, with no less than 2 hidden layers and no less than 20 neurons in each layer.

[0040] The parameters of the neural network structure are randomly initialized. The neural network is trained using the dataset described in step three. The dataset is randomly divided into three parts: a training set, a test set, and a validation set. Training, testing, and validation are performed on each part separately. The performance metrics of the neural network, such as mean squared error, are optimized using a gradient descent algorithm until the metrics are less than a threshold ε, at which point training ends. The trained neural network can output the switching time of each nozzle's RCS based on the input impulse torque command and nozzle health status.

[0041] t rcs =f NN (H cmd ,g rcs (13)

[0042] Among them, f NN It is a neural network-based RCS intelligent fault-tolerant allocator.

[0043] Step 5: Apply the distributor to actual aircraft attitude control tasks to efficiently generate the on / off time commands of each nozzle in the RCS under normal or fault conditions, ensuring that the RCS output can meet the control command requirements and realize fault-tolerant distribution control of the aircraft attitude.

[0044] Beneficial effects:

[0045] 1. The present invention discloses a neural network-based intelligent RCS fault-tolerant allocation method for aircraft. By integrating highly practical tracking task instructions and random instructions covering the entire RCS operation space, the generated dataset can effectively train an intelligent RCS allocator with both high precision and strong generalization. The fault-tolerant allocation capability of aircraft attitude control is enhanced through the intelligent RCS fault-tolerant allocator.

[0046] 2. The present invention discloses an intelligent fault-tolerant allocation method for RCS of aircraft based on neural networks. The nozzle health status is used as one of the inputs, and fault modes are systematically introduced so that the neural network can adaptively learn the allocation strategy under various fault modes and has the ability to allocate faults under RCS faults.

[0047] 3. The present invention discloses an intelligent fault-tolerant allocation method for RCS of aircraft based on neural networks. It uses neural networks to calculate the nozzle opening time, does not rely on optimization algorithms for allocation, has low requirements for equipment computing power, and ensures the real-time performance of allocation.

[0048] 4. The present invention discloses an intelligent fault-tolerant allocation method for RCS of aircraft based on neural networks. It establishes an RCS model through a control efficiency matrix, transforms the RCS control allocation problem into an optimization problem to solve and generate a dataset. It does not have strict requirements on the specific number and installation position of RCS nozzles, and has a wider range of applications and higher flexibility compared with table lookup allocation methods. Attached Figure Description

[0049] Figure 1 This is a structural block diagram of a neural network-based intelligent fault-tolerant allocation method for aircraft RCS (Rapid Cross Section) disclosed in this invention.

[0050] Figure 2 This describes the change in attitude angle tracking error of the present invention under fault-free conditions.

[0051] Figure 3 This describes the RCS allocation scenario under fault-free conditions.

[0052] Figure 4 The attitude angle tracking error of the present invention changes under the scenario of nozzle 2 failure;

[0053] Figure 5 This describes the RCS fault tolerance allocation of the present invention under the scenario of nozzle 2 failure. Detailed Implementation

[0054] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0055] Example 1:

[0056] Considering scenarios where aircraft perform cross-domain maneuvers in space and fly in the upper atmosphere, due to the thinness and low dynamic pressure of the upper atmosphere, the control capability of aerodynamic control surfaces is weak. The aircraft mainly relies on six RCS propulsion nozzles installed at different positions in the plane for attitude control.

[0057] like Figure 1 As shown in the figure, this embodiment discloses a neural network-based intelligent fault-tolerant allocation method for aircraft RCS, and the specific implementation steps are as follows:

[0058] Step 1: Establish the three-degree-of-freedom attitude dynamics equations for the aircraft.

[0059] The three-degree-of-freedom attitude dynamics equations of the aircraft during flight are as follows:

[0060]

[0061] in, Let γ be the Euler angle vector of the aircraft. ψ represents the roll angle, pitch angle, and yaw angle of the aircraft, respectively; ω = [ω x ,ω y ,ω z ] T Let ω be the three-axis attitude angular velocity vector of the aircraft. × I is the skew-symmetric matrix corresponding to ω; I is the aircraft's moment of inertia matrix; R is the transformation matrix, specifically expressed as follows:

[0062]

[0063] M is the torque acting on the aircraft, specifically expressed as follows:

[0064] M = M base +M ctrl +f (16)

[0065] Among them, M base The fundamental torques of an aircraft include the fundamental aerodynamic torque of the fuselage, the gravity gradient torque, etc.; M ctrl The control torque generated by the aircraft's control actuators; f represents various disturbances and uncertainties. For an aircraft that uses only RCS for attitude control, the thrust and torque generated when each nozzle of the RCS is open are fixed, and the aircraft's control torque is the sum of the torques output by each nozzle:

[0066]

[0067] Where n is the number of nozzles, in this embodiment, n = 6; M ctrl,i M represents the torque output by the i-th nozzle (i = 1, 2, ..., n); ri Let be the torque generated when the i-th nozzle is opened.

[0068] Step 2: Based on the three-degree-of-freedom attitude dynamics equations of the aircraft, establish the RCS control allocation model of the aircraft.

[0069] An RCS system consists of a series of thrust nozzles. It generates a reverse thrust force that does not pass through the spacecraft's center of mass by ejecting gas in a specified direction around the periphery, thereby producing a control torque. For an RCS system with n nozzles, the control efficiency matrix of the RCS system is defined as follows:

[0070] B rcs =[M r1 M r2 ,...,M rn (18)

[0071] Considering the discrete torque output characteristics of RCS, the allocation essentially involves calculating the opening time of each nozzle within a specified time period Δt, so that the sum of the impulse torques generated by each nozzle approaches the control command torque M. cmd The corresponding impulse torque command H cmd =M cmd Δt, that is:

[0072] B rcs t rcs →H cmd (19)

[0073] Among them, t rcs =[t r1 ,t r2 ,...,t rn ] T This indicates the opening time of each nozzle.

[0074] Step 3: Transform the aircraft RCS control allocation problem into an optimization problem, solve the RCS control allocation problem, and design and generate a dataset for training the neural network.

[0075] RCS control allocation can be transformed into an optimization problem, which can optionally be solved using optimization methods such as nonnegative least squares or linear programming. In this embodiment, considering the nonnegativity of the nozzle opening time, the RCS control allocation problem is transformed into the following optimization problem:

[0076] Therefore, given B rcs and H cmdAt that time, the non-negative least squares method is used to solve for the opening time of each nozzle, and a dataset sample is generated.

[0077] When generating the dataset for training the RCS intelligent fault-tolerant assignment neural network, a hybrid strategy is employed to generate the impulse moment instruction set {H}. cmd This strategy combines two methods: generating instructions that track the nominal trajectory and generating random instructions that cover the entire RCS operating space.

[0078] {H cmd}={H cmd , track}∪{H cmd , rand} (twenty one)

[0079] The command to track the nominal trajectory originates from the impulse moment command H generated by the aircraft tracking the reference attitude profile during the actual mission. cmd,track Specifically, it is described as follows:

[0080] H cmd , track =f(Ω) ref ,ω ref ,ΔΩ init ,d) (22)

[0081] Among them, Ω ref and ω ref These are the reference Euler angles and angular velocity, ΔΩ, respectively. init Let be the initial attitude deviation, and d represent external disturbances and system uncertainties. These instructions form the foundation of the dataset, ensuring the high applicability of the strategies learned by the neural network in real-world tasks.

[0082] Random instruction generation covering the entire RCS operating space refers to instruction generation based on the control efficiency matrix B. rcs Determine the boundary of the impulse moment that the RCS can generate within a specified time period Δt [H] min H max Within this boundary, impulse moment command H is generated by uniformly and randomly sampling. cmd,rand :

[0083] H cmd,rand ~U(H min H max (23) Among them,

[0084] This allows for a comprehensive exploration of the capabilities of the RCS system, covering extreme cases and enhancing the generalization ability of the dataset. In this embodiment, the ratio of the instruction dataset tracking the nominal trajectory to the random instruction dataset covering the entire RCS operating space is 0.7:0.3.

[0085] The input to the dataset samples for designing the RCS intelligent fault-tolerant assignment neural network includes a 3D impulse moment command H. cmd and n-dimensional nozzle health status:

[0086] g rcs =[g r1 ,g r2 ,...,g ri ,...,g rn ] T , g ri ∈[0,1], i=1,2,...,n (25)

[0087] Among them, g ri A value of 1 indicates that the i-th nozzle is working normally; g ri A value of 0 indicates a complete failure of the i-th nozzle, resulting in zero output torque, corresponding to matrix B. rcs Set the i-th column of g to zero; ri The values ​​between 0 and 1 represent the degree of decrease in nozzle efficiency and output torque; the output of the dataset samples is the solved nozzle opening time t. rcs ;

[0088] When generating the dataset, nozzle damage is considered, i.e., g is modified. rcs The values ​​of the elements in the middle characterize single nozzle failure, multi-nozzle failure, and different degrees of nozzle efficiency degradation, ensuring that the neural network can learn allocation strategies under different failure conditions and enhance fault-tolerant allocation capabilities.

[0089] Step 4: Design a neural network. Train the neural network using the dataset obtained in Step 3 to obtain the RCS intelligent fault-tolerant allocator.

[0090] Design the structure of the neural network, with the input layer including a 3-dimensional impulse moment instruction H. cmd and n-dimensional nozzle health status g rcs The output layer includes the opening time t of each nozzle. rcs Optionally, a fully connected neural network structure is adopted, with no less than 2 hidden layers and no less than 20 neurons in each layer.

[0091] The parameters of the neural network structure are randomly initialized. The neural network is trained using the dataset described in step three. The dataset is randomly divided into three parts: a training set, a test set, and a validation set. Training, testing, and validation are performed on each part separately. The performance metrics of the neural network, such as mean squared error, are optimized using a gradient descent algorithm until the metric is less than a threshold ε, at which point training ends. In this embodiment, the number of hidden layers is 3, and each layer has 80 neurons. The ratio of the training set, test set, and validation set is 0.8:0.1:0.1. The trained neural network can output the on / off time of each nozzle's RCS based on the input impulse torque command and nozzle health status.

[0092] t rcs =f NN (H cmd ,g rcs (26)

[0093] Among them, f NN It is a neural network-based RCS intelligent fault-tolerant allocator.

[0094] Step 5: Apply the distributor to the aircraft attitude control task in this embodiment to efficiently generate the switching time commands of each nozzle in the RCS under normal or fault conditions, ensuring that the RCS output can meet the control command requirements and realize the fault-tolerant distribution control of the aircraft attitude.

[0095] In this embodiment, a neural network-based RCS intelligent fault-tolerant allocator is designed and trained using the method of this invention. It is then applied to a typical aircraft attitude control task, yielding the attitude angle tracking error variation of this invention during a flight under a fault-free scenario, as shown below. Figure 2 As shown, the three attitude angle errors converge rapidly from their initial large deviations to near zero. Due to the discrete characteristics and disturbances in the RCS torque output, the errors fluctuate within a small range. The RCS allocation of this invention under a fault-free scenario is as follows. Figure 3 As shown, each nozzle can function normally and is activated at specific times to meet attitude control requirements.

[0096] Considering the scenario where nozzle 2 malfunctions and cannot output reaction torque, the attitude angle tracking error change of this invention during a certain flight is as follows: Figure 4 As shown, under fault conditions, the errors of the three attitude angles can still quickly converge from the initial large deviation to near 0. Due to the discrete characteristics and disturbances of the RCS torque output, the error will fluctuate within a small range. The RCS fault tolerance allocation of this invention under the nozzle 2 fault scenario is as follows. Figure 5 As shown, when the output of nozzle 2 is 0, the intelligent fault-tolerant distributor actively calls on other nozzles to participate in the distribution, reliably completing the attitude control task, demonstrating that the distributor has strong fault tolerance capability.

[0097] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is 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. A neural network-based intelligent fault-tolerant allocation method for aircraft RCS, characterized in that: Includes the following steps: Step 1: Establish the three-degree-of-freedom attitude dynamics equations of the aircraft; Step 2: Based on the three-degree-of-freedom attitude dynamics equations of the aircraft, establish the RCS control allocation model of the aircraft; Step 3: Transform the aircraft RCS control allocation problem into an optimization problem, solve the RCS control allocation problem, and design and generate a dataset for training the neural network; Step 4: Construct a neural network. Train the neural network using the dataset obtained in Step 3 to obtain the RCS intelligent fault-tolerant allocator, which enhances the fault-tolerant allocation capability for aircraft attitude control.

2. The method as described in claim 1, characterized in that: It also includes step five, which applies the RCS intelligent fault-tolerant allocator obtained in step four to the actual aircraft attitude control task, generates the switching time commands of each nozzle of the RCS, ensures that the RCS output can meet the control command requirements, and realizes the fault-tolerant allocation control of the aircraft attitude.

3. The method as described in claim 1, characterized in that: The three-degree-of-freedom attitude dynamics equations of the aircraft in step one are expressed as follows: in, Let γ be the Euler angle vector of the aircraft. ψ represents the roll angle, pitch angle, and yaw angle of the aircraft, respectively; ω = [ω x ,ω y ,ω z ] T Let ω be the three-axis attitude angular velocity vector of the aircraft. × I is the skew-symmetric matrix corresponding to ω; I is the aircraft's moment of inertia matrix; R is the transformation matrix, specifically expressed as follows: M is the torque acting on the aircraft, specifically expressed as follows: M=M base +M ctrl +f (3) Among them, M base The fundamental torques of an aircraft include the fundamental aerodynamic torque of the fuselage, the gravity gradient torque, etc.; M ctrl The control torque generated by the aircraft's control actuators; f represents various disturbances and uncertainties; for an aircraft that uses only RCS for attitude control, the thrust and torque generated when each nozzle of the RCS is open are fixed, and the aircraft's control torque is the sum of the torques output by each nozzle: Where n is the number of nozzles; M ctrl,i M represents the torque output by the i-th nozzle (i = 1, 2, ..., n); ri Let be the torque generated when the i-th nozzle is opened.

4. The method as described in claim 1, characterized in that: The implementation method for step two is as follows: For an RCS system with n nozzles, the control efficiency matrix of the RCS system is defined as follows: B rcs =[M r1 ,M r2 ,...,M ri ,...,M rn ],i=1,2,...,n (5) Considering the discrete torque output characteristics of RCS, the allocation essentially involves calculating the opening time of each nozzle within a specified time period Δt, so that the sum of the impulse torques generated by each nozzle approaches the control command torque M. cmd The corresponding impulse torque command H cmd =M cmd Δt, that is: B rcs t rcs →H cmd (6) Among them, t rcs =[t r1 ,t r2 ,...,t rn ] T This indicates the opening time of each nozzle.

5. The method as described in claim 1, characterized in that: The implementation method for step three is as follows: Considering the non-negativity of nozzle opening time, the RCS control allocation problem is transformed into the following optimization problem: Therefore, given B rcs and H cmd At that time, the non-negative least squares method was used to solve for the opening time of each nozzle, and a dataset sample was generated; When generating the dataset for training the RCS intelligent fault-tolerant assignment neural network, a hybrid strategy is employed to generate the impulse moment instruction set {H}. cmd This strategy combines two methods: generating instructions that track the nominal trajectory and generating random instructions that cover the entire RCS operating space. {H cmd }={H cmd,track }∪{H cmd, r and } (8) The command to track the nominal trajectory originates from the impulse moment command H generated by the aircraft tracking the reference attitude profile during the actual mission. cmd,track ,Right now: H cmd,track =f(Ω ref ,oh ref ,LOOK init ,d) (9) Among them, Ω ref and ω ref These are the reference Euler angles and angular velocity, ΔΩ, respectively. init Let d represent the initial attitude deviation, and d represent external disturbances and system uncertainties. Random instruction generation covering the entire RCS operating space refers to instruction generation based on the control efficiency matrix B. rcs Determine the boundary of the impulse moment that the RCS can generate within a specified time period Δt [H] min H max Within this boundary, impulse moment command H is generated by uniformly and randomly sampling. cmd,rand : H cmd,rand ~U(H min ,H max ) (10) in, 6. The method as described in claim 1, 2, 3, 4 or 5, characterized in that: When generating the dataset for training the neural network as described in step three, the nozzle damage situation also needs to be considered; The input to the dataset samples includes a 3D impulse moment command H. cmd and n-dimensional nozzle health status: g rcs =[g r1 ,g r2 ,...,g ri ,...,g rn ] T ,g ri ∈[0,1],i=1,2,...,n (12) Among them, g ri A value of 1 indicates that the i-th nozzle is working normally; g ri A value of 0 indicates a complete failure of the i-th nozzle, resulting in zero output torque, corresponding to matrix B. rcs Set the i-th column to zero; g ri The values ​​between 0 and 1 represent the degree of decrease in nozzle efficiency and output torque; the output of the dataset samples is the solved nozzle opening time t. rcs ; When generating the dataset, nozzle damage is considered, i.e., g is modified. rcs The values ​​of the elements in the middle characterize single nozzle failure, multi-nozzle failure, and different degrees of nozzle efficiency degradation, ensuring that the neural network can learn allocation strategies under different failure conditions and enhance fault-tolerant allocation capabilities.

7. The method as described in claim 1, characterized in that: The neural network structure described in step four is such that the input layer includes a 3-dimensional impulse moment instruction H. cmd and n-dimensional nozzle health status g rcs The output layer includes the opening time t of each nozzle. rcs .

8. The method as described in claim 7, characterized in that: A fully connected neural network structure is adopted, with no fewer than 2 hidden layers and no fewer than 20 neurons in each layer.

9. The method as described in claim 1, characterized in that: Step four describes the method of training a neural network using the dataset obtained in step three to obtain the RCS intelligent fault-tolerant allocator. The parameters of the neural network structure are randomly initialized. The neural network is trained using the dataset obtained in step three. The dataset is randomly divided into three parts: a training set, a test set, and a validation set, which are then used for training, testing, and validation respectively. The performance index of the neural network is optimized using the gradient descent algorithm until the index is less than a threshold ε, at which point training ends. The trained neural network can output the switching time of each nozzle in the RCS based on the input impulse torque command and nozzle health status. t rcs =f NN (H cmd ,g rcs ) (13) Among them, f NN It is a neural network-based RCS intelligent fault-tolerant allocator.

Images