Multi-form carrier rocket simulation method, system and device based on LTSM neural network and medium

Abstract

The invention belongs to the technical field of aerospace, and particularly relates to a multi-form carrier rocket simulation method, system and device based on an LTSM neural network and a medium. The simulation method comprises the steps of firstly setting an initialization variable and an initialization environment parameter of a carrier rocket, then optimizing a learning rate and a gradient clipping rate of a built LSTM neural network by adopting a loss function optimization algorithm, and then training the optimized LSTM neural network by adopting an Adam optimizer. Inputting the dynamic variable and the environment variable into the trained LSTM neural network to simulate the attitude adjustment force and the thrust moment of the carrier rocket, and finally obtaining a flight simulation result of the carrier rocket; the simulation system comprises an initial parameter acquisition module, an LSTM neural network training module, an instantaneous parameter acquisition module and a rocket flight simulation module. According to the invention, the accuracy and reliability of rocket flight simulation can be greatly improved, so that the rocket can keep a stable flight state when facing a complex and variable external interference environment.

Description

Technical Field

[0001] The present invention belongs to the field of aerospace technology, and in particular relates to a multi-modal launch vehicle simulation method, system, equipment and medium based on LTSM neural network. Background Art

[0002] Rocket launch involves a complex multi-physical system, including dynamics, aerodynamics, propulsion, fuel consumption and other factors. The rocket propulsion system needs to adjust the thrust according to the needs of each stage of the flight process to overcome the effects of gravity, air resistance and other factors to achieve the expected flight trajectory. In the field of rocket flight simulation, there are still more traditional simulation methods such as PID control algorithms. Such simulation methods are not real-time and lack predictability. Due to the problem of preset parameters, the predicted trajectory is not ideal and cannot accurately simulate the actual rocket launch. Specifically, in the existing rocket launch system, control and trajectory optimization mainly rely on rule-oriented control systems and simple feedback mechanisms, but these methods have the following major defects:

[0003] 1. Lack of adaptability:

[0004] Traditional rocket launch control systems often adjust thrust and combustion rate based on predetermined parameters and fixed control logic. These preset rules cannot be adaptively adjusted when faced with rapid changes in the external environment (such as drastic changes in disturbances such as air pressure and wind speed), resulting in insufficient flexibility and responsiveness of the system.

[0005] 2. Limitations of real-time optimization:

[0006] Existing control methods are difficult to perform real-time global optimization based on the constantly changing parameters during flight. For thrust distribution, attitude control, etc., existing control methods (such as PID control) cannot respond to complex disturbances quickly and effectively.

[0007] 3. Poor posture control effect:

[0008] Under complex aerodynamic conditions, the rocket's attitude control needs to deal with multiple factors such as lateral wind speed and turbulence. Traditional control systems usually rely on simple feedback mechanisms, such as offsetting lateral wind speed through lateral thrust. This method cannot ensure the rocket's stable attitude under multiple disturbances.

[0009] In the prior art, the patent publication number is "CN118518102A" and the name is "A method for online planning of flight trajectories based on parameter identification". A method for online planning of flight trajectories based on parameter identification is disclosed. By establishing a deviation identification model of each trajectory sensitive parameter, an online identification method for trajectory sensitive parameter deviation is adopted. Based on the sensor measurement data during the flight, the online deviation identification of each parameter is realized. Based on the deviation of the trajectory sensitive parameters, the corresponding flight trajectory is selected from the trajectory library to complete the online planning of the flight trajectory. However, this method is more suitable for aircraft with non-adjustable thrust or no thrust, and does not have obvious advantages for aircraft with adjustable thrust. In addition, if the dynamic characteristics of the aircraft deviate greatly from the model, the planning effect may be affected. At the same time, for scenarios where the flight environment changes rapidly (such as strong winds and turbulence), the trajectory library of this method may not be sufficient to cope with dynamic and complex environmental conditions, affecting the robustness of the system.

[0010] In the prior art, the patent publication number is "CN 109343341A" and the name is "A method for intelligent control of vertical recovery of launch vehicles based on deep reinforcement learning". It discloses a method for intelligent control of vertical recovery of launch vehicles based on deep reinforcement learning. By establishing a simulation model of vertical recovery of launch vehicles, establishing a Markov decision process, and using a neural network model, a method for autonomous intelligent control of launch vehicles is realized. It has the effects of vertical recovery attitude control and trajectory planning of launch vehicles, but it cannot intelligently control thrust, and the neural network model does not consider the learning rate return problem.

[0011] In the prior art, the patent publication number is "CN 116697829 A" and the name is "A rocket landing guidance method and system based on deep reinforcement learning". It discloses a rocket landing guidance method and system based on deep reinforcement learning. By building a rocket six-degree-of-freedom dynamics model simulation environment, establishing a Markov decision process, building a neural network according to the deep reinforcement learning algorithm, and using the trained neural network model to guide the rocket landing flight, the function of the landing flight of a recoverable carrier rocket is realized. It has the effects of high algorithm efficiency, low fuel consumption, and good landing accuracy, but it has the problem of insufficient environmental influencing factors. Summary of the invention

[0012] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to propose a multi-modal launch vehicle simulation method, system, equipment and medium based on the LTSM neural network. The method builds a neural network through a deep learning algorithm, trains the built neural network, and then calls the trained neural network for simulation verification, thereby realizing the trajectory optimization of the launch vehicle during flight and the autonomous control capability after being disturbed, and can perform high-precision simulation of the rocket flight trajectory and thrust control in the presence of multiple environmental disturbance factors.

[0013] In a first aspect, a multi-modal launch vehicle simulation method based on a LTSM neural network comprises the following steps:

[0014] S1: Set the initialization variables of the launch vehicle, including the initial mass m of the launch vehicle 0 , Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Total mass of the rocket m = m 0 +m 1 +m Fuel ;

[0015] S2: Setting the initialization environment parameters of the launch vehicle, which include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial atmospheric temperature T 0 , air resistance F drag , Turbulence I ` , initial atmospheric density ρ 0 ;

[0016] S3: Building an LSTM (Long Short-Term Memory) neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, a LSTM output layer dimension, and an activation function are set;

[0017] S4: using a loss function optimization algorithm to optimize the learning rate and gradient clipping rate of the LSTM neural network in step S3 to obtain an optimized LSTM neural network, and then using an Adam (adaptive moment estimation) optimizer to train the optimized LSTM neural network to obtain a trained LSTM neural network;

[0018] S5: Initialize the variables in step S1 and solve the dynamic variables of the launch vehicle through the dynamic equation. The dynamic variables of the launch vehicle include the total thrust F, acceleration The time-varying mass m(t), the derivative of the height h, dh, and gravity Attitude adjustment forces and launch vehicle thrust vector control torque;

[0019] S6: according to the environmental parameters initialized in step S2, using the environmental parameter expression to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, wind speed U;

[0020] S7: The dynamic variables and environmental variables calculated in step S5 and step S6 are input into the LSTM neural network trained in step S4 to simulate the attitude adjustment force and thrust torque of the launch vehicle, and obtain the flight simulation result of the launch vehicle.

[0021] Furthermore, the LSTM neural network construction process in step S3 is as follows:

[0022] S3.1: Obtain time series data from the initialization variables described in step S1 and the initialization environmental parameters described in step S2, including time t (current time), speed v (current speed), initial atmospheric density ρ 0 , Single engine thrust Initial fuel mass m Fuel ;

[0023] S3.2: Normalize the time series data in step S3.1, and normalize the data x norm As shown in formula (1):

[0024]

[0025] In formula (1), x is the unnormalized data, μ is the mean, and σ is the standard deviation;

[0026] S3.3: Use the normalized data described in step S3.2 to build the LSTM neural network described in step S3, including an input gate, a forget gate, and an output gate.

[0027] Furthermore, the parameters in the loss function optimization algorithm in step S4 are as follows:

[0028] The mean square error (MSE) is shown in formula (2):

[0029]

[0030] In formula (2), N is the total number of data, y true,t is the true value at time t, y pred,t is the predicted value at time t;

[0031] The angle change penalty term (change_penalty) is shown in formula (3):

[0032]

[0033] In formula (3), (y pred,t+1 -y pred,t ) is the change in the predicted value between time t and time t+1; (y true,t+1 -y true,t ) is the change in true value between time t and time t+1;

[0034] The angular velocity constraint (angle change rate) is shown in formula (4):

[0035]

[0036] If the angle change rate exceeds a set maximum value, an angle rate penalty is applied, where rate penalty = max(0, angle change rate - maxangle change)

[0037] The total loss function (Total Loss) is shown in formula (5):

[0038] Total Loss = MSE + λ 1 change_penalty+λ 2 ·Rate penalty (5)

[0039] In formula (5), λ1 and λ2 are the weight coefficients of change_penalty and rate_penalty, respectively, which control the influence of the penalty term.

[0040] Furthermore, the momentum update, second-order moment estimation, and parameter update in the Adam optimizer in step S4 are respectively shown in formulas (6), (7), and (8):

[0041] m t =β 1 m t-1 +(1-β 1 ) t (6)

[0042] In formula (6), m t is the momentum estimate, which indicates the cumulative information of the direction and magnitude of the gradient, m t-1 is the first moment estimate of the gradient (momentum), β 1 is the decay rate of the first-order moment, g t is the gradient at the current time;

[0043]

[0044] In formula (7), v t is the second-order moment estimate, which represents the weighted sum of squares of the gradient magnitude and is used to adjust the update step size, v t-1 is the second-order moment estimate of the gradient, β 2 is the decay rate of the second-order moment;

[0045]

[0046] In formula (8), θ t is the model parameter at the current time step, θ t+1 is the model parameter for the next time step, α is the global learning rate, which controls the step size of parameter update, and ∈ is a smoothing term used to avoid division by zero errors.

[0047] Furthermore, the expression of the kinetic equation in step S5 is shown in formula (9):

[0048]

[0049] In formula (9), m is the total mass of the rocket, m(t) is the sum of the mass of the rocket and the mass of the fuel at time t, and the thrust Air resistance gravity The total thrust F formed by the random disturbance determines the acceleration is the instantaneous combustion rate of the fuel; v e is the effective exhaust velocity, c e is the fuel consumption efficiency coefficient; is the rocket acceleration, F is the total thrust, v is the velocity, γ is the attitude angle, and h is the altitude.

[0050] Furthermore, the expressions of the attitude adjustment force and the launch vehicle thrust vector control torque in step S5 are shown in formula (10):

[0051]

[0052] In formula (10), is the posture adjustment force, K p is the proportional gain coefficient, controlling the thrust compensation for the lateral wind speed, v wind is the wind speed, M tvc is the thrust vector control torque of the launch vehicle, and Δθ is the attitude adjustment angle.

[0053] Furthermore, the environmental parameter expression in step S6 is shown in formula (11):

[0054]

[0055] In formula (11), ρ 0 , P 0 They represent the atmospheric density and atmospheric pressure at the surface respectively; k is the von Karman constant, u * is the friction velocity, U(z) is the average wind speed at height z; Γ is the vertical temperature gradient; T 0 is the sea level temperature; the temperature at the top of the troposphere (about 11 km) becomes constant; e(z) is the water vapor pressure at height z; e0 is the water vapor pressure at sea level; M v is the molar mass of water vapor; R v is the gas constant of water vapor; T is the atmospheric temperature, T 0 is the initial atmospheric temperature, Γh is the product of the vertical temperature lapse rate and the variable, M v is the molar mass of water vapor, R v is the gas constant for water vapor.

[0056] In the second aspect, a multi-form launch vehicle simulation system based on an LTSM neural network is provided, and the multi-form launch vehicle simulation method is applied. The simulation system comprises an initial parameter acquisition module, an LSTM neural network training module, an instantaneous parameter acquisition module and a rocket flight simulation module:

[0057] Initial parameter acquisition module: set the initial variables of the launch vehicle, including the initial mass m of the launch vehicle 0 , Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Set the initialization environment parameters of the launch vehicle, the initialization environment parameters include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial temperature T 0 , air resistance F drag , turbulence disturbance I ` , initial atmospheric density ρ 0 ; Total mass of the rocket m = m 0 +m 1 +m Fuel ;

[0058] LSTM neural network training module: build an LSTM neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, an LSTM output layer dimension and an activation function are set; a loss function optimization algorithm is used to optimize the learning rate and the gradient clipping rate of the LSTM neural network to obtain an optimized LSTM neural network, and then an Adam optimizer is used to train the optimized LSTM neural network to obtain a trained LSTM neural network;

[0059] Instantaneous parameter acquisition module: According to the initialization variables, the dynamic variables of the launch vehicle are solved through the dynamic equation. The dynamic variables of the launch vehicle include total thrust F, acceleration The time-varying mass m(t), the derivative dh of the height h, and the gravity attitude adjustment force and launch vehicle thrust vector control torque; according to the initialized environmental parameters, using environmental parameter expressions to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, wind speed U;

[0060] Rocket flight simulation module: The dynamic variables and environmental variables are input into the trained LSTM neural network to simulate the attitude adjustment force and thrust torque of the launch vehicle to obtain the launch vehicle flight simulation result.

[0061] In a third aspect, an electronic device includes a memory and a processor:

[0062] Memory: used for storing a computer program for implementing the multi-modal launch vehicle simulation method based on the LTSM neural network;

[0063] Processor: used to implement the multi-modal launch vehicle simulation method based on LTSM neural network when executing the computer program.

[0064] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-modal launch vehicle simulation method based on the LTSM neural network is implemented.

[0065] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0066] 1. The present invention builds an LSTM neural network through step 3, inputs environmental parameters (wind speed, atmospheric pressure, atmospheric density, temperature, water vapor pressure, etc.) into the LSTM neural network, adjusts the rocket direction angle, thrust and speed in real time, and performs secondary learning through negative feedback and stores them in long-term memory;

[0067] 2. The present invention optimizes the LSTM neural network through the loss function optimization algorithm of step 4, trains the LSTM neural network through the Adam optimizer, and improves the flight stability of the carrier rocket. Specifically, the gradient clipping technology and the learning rate adaptive adjustment of the LSTM neural network in step 4 are used, so that in extreme cases, the optimizer can still effectively optimize and adjust the model parameters, and the thrust distribution strategy can be continuously improved and optimized during the flight, effectively avoiding the instability problem in the neural network training. Especially in the case of rapid changes in thrust, the adaptive adjustment of the LSTM neural network makes the training process more stable, reduces the risk of model non-convergence, enables the rocket to learn and adapt to different flight environments and states (including high dynamic and complex flight scenes), avoids the phenomenon of gradient explosion or disappearance, and further improves the accuracy of thrust compensation.

[0068] 3. The present invention acquires environmental data in real time through steps 1 to 4 and performs LSTM neural network training on the rocket, and performs rocket flight simulation according to the real-time data collected and calculated in steps 5 to 7, and can simulate environmental parameters such as air pressure, wind speed, temperature, humidity and turbulence, and make corresponding simulation adjustments to the rocket attitude and thrust torque. Specifically, step 7 controls the attitude of the rocket by the linkage of the collection of environmental data and the thrust distribution model, and adjusts the thrust distribution in real time. This thrust control method based on multidimensional data improves the flight stability of the rocket, so that the rocket can still maintain a stable flight state when facing a complex and changeable external interference environment.

[0069] In summary, the present invention utilizes the trained LSTM neural network and integrates multiple physical field effects such as aerodynamics, gravity, combustion chemistry and environmental disturbances in rocket flight, and utilizes multi-equation coupling to simulate the dynamic characteristics of the rocket in various flight stages. The LSTM neural network is used to train the optimal parameters in different environments during rocket flight, which greatly improves the accuracy and reliability of the simulation. It can more accurately display the trajectory optimization and prediction during the rocket flight, and adjust the thrust at the same time to ensure that the rocket still flies according to the predetermined trajectory when it receives disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The present invention provides a flow chart of a multi-modal launch vehicle simulation method based on LTSM neural network.

[0071] Figure 2 This is a flow chart of building an LSTM neural network provided by the present invention.

[0072] Figure 3 It is a schematic diagram of the rocket configuration provided by an embodiment of the present invention.

[0073] Figure 4 This is a comparison chart of the verification loss rate and training loss rate of the LSTM neural network provided by an embodiment of the present invention and other models.

[0074] Figure 5 It is a height change curve simulated by using LSTM neural network provided by an embodiment of the present invention.

[0075] Figure 6 This is a height change curve provided by an embodiment of the present invention without using LSTM neural network simulation.

[0076] Figure 7 This is a fuel consumption comparison chart before and after the use of the LSTM neural network provided by an embodiment of the present invention.

[0077] Figure 8This is a speed comparison diagram before and after the LSTM neural network provided by an embodiment of the present invention is used.

[0078] Fig. 9 This is a posture control angle change diagram provided by an embodiment of the present invention without using an LSTM neural network.

[0079] Fig.10 This is a graph of posture control angle changes after using an LSTM neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0080] Combine the following Figures 1 to 10 The present invention is further described in detail with reference to the embodiments. The launch vehicle described in the present invention refers only to a two-stage rocket and a three-stage rocket. Figure 1 The flowchart of the multi-modal launch vehicle simulation method based on the LTSM neural network is shown in FIG. 1 . The rocket simulation method includes the following steps:

[0081] S1: Set the initialization variables of the launch vehicle, including the initial mass m of the launch vehicle 0 , Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Total mass of the rocket m = m 0 +m 1 +m Fuel ;

[0082] This embodiment uses the Python compiler to implement neural network learning and rocket flight simulation functions, where the initial mass of the launch vehicle is m 0 is 10000kg, load m 1 500kg, starting launch angle Altitude h = 0m, speed v = 0m, time t = 0s, initial fuel mass m Fuel =8000kg, initial fuel burning rate α = 50kg / s, single engine thrust Configuration n = 4 + 1; Figure 3 The schematic diagram of the rocket configuration of this embodiment is shown, the configuration n=4+1, that is, four rocket engines surround one rocket engine to provide propulsion power;

[0083] S2: Setting the initialization environment parameters of the launch vehicle, which include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial temperature T 0 , air resistance F drag , Turbulence I `, initial atmospheric density ρ 0 ;

[0084] In this embodiment, the turbulence parameter I=5, the initial air pressure P 0 =101.325kPa, wind speed v wind =3m / s, initial temperature T 0 =288.15K, air resistance F drag =10N, turbulence disturbance I ` =1, initial atmospheric density ρ 0 =1.225kg / m 3 ;

[0085] S3: Building an LSTM neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, a LSTM output layer dimension, and an activation function are set;

[0086] The LSTM neural network construction process in step S3 is as follows:

[0087] S3.1: Obtain time series data from the initialization variables described in step S1 and the initialization environmental parameters described in step S2, including time t (current time), speed v (current speed), initial atmospheric density ρ 0 , Single engine thrust Initial fuel mass m Fuel ;

[0088] S3.2: Normalize the time series data in step S3.1, and normalize the data x norm As shown in formula (1):

[0089]

[0090] In formula (1), x is the unnormalized data, μ is the mean, and σ is the standard deviation;

[0091] S3.3: Use the normalized data in step S3.2 to build the LSTM neural network in step S3, including an input gate, a forget gate, and an output gate;

[0092] The input gate, forget gate, and output gate process data in sequence, where the input gate is used to control whether information enters the memory unit, the forget gate is used to determine whether the information in the memory unit is forgotten, and the output gate is used to control whether the information in the memory unit is output; the memory unit is updated through the combination of the input gate and the forget gate;

[0093] Input gate vector i t As shown in formula (2):

[0094] i t=σ(W i ·[h t-1 ,x t ]+b i ) (2)

[0095] In formula (2), i t is the input gate vector, which controls the influence of the time series data in step S3.1 on the current memory unit, and the element value is between [0,1]; σ is the Sigmoid activation function, which maps the time series data to the interval [0,1]; h t-1 is the hidden state vector of the previous moment, which contains the output information of the previous moment; x t is the input vector at the current moment, containing the observation data at the current moment; W i is the weight matrix of the input gate; [h t-1 ,x t ] is the hidden state vector h of the previous moment t-1 and the input vector x at the current moment t The vector formed by series connection (concatenation) is used to multiply the weight matrix of each gate; b i is the bias vector of the input gate;

[0096] Forget gate vector f t As shown in formula (3):

[0097] f t =σ(W f ·[h t-1 ,x t ]+b f ) (3)

[0098] In formula (3), f t is the forget gate vector, which determines the degree of forgetting of information in the memory unit, and the element value is between [0,1]; W f is the weight matrix of the forget gate; b f is the bias vector of the forget gate;

[0099] Output gate vector o t As shown in formula (4):

[0100] o t =σ(W o ·[h t-1 ,x t ]+b o ) (4)

[0101] In formula (4), o t is the output gate vector, which controls how much information in the memory cell is output to the hidden state h t , element values ​​are between [0,1]; W ois the weight matrix of the output gate; b o is the bias vector of the output gate;

[0102] The memory cell state vector c at the current moment t As shown in formula (5):

[0103] c t =f t ⊙c t-1 +i t ⊙tanh(W c ·[h t-1 ,x t ]+b c ) (5)

[0104] In formula (5), c t is the memory unit state vector at the current moment, which accumulates important information up to the current moment; c t-1 is the memory unit state vector at the previous moment; W c is the weight matrix used to generate candidate memories; b c is the bias vector used to generate candidate memories;

[0105] The hidden state of LSTM involves the hidden state vector h at the current moment t As shown in formula (6):

[0106] h t =o t ⊙tanh(c t )(6)

[0107] In formula (6), h t is the hidden state vector at the current moment, which is used as the output at the current moment and is also used for the calculation at the next moment; o t is the output gate vector; c t is the memory unit state vector at the current moment.

[0108] Among the above parameters, W i , W f , W o , W c is the weight matrix of each gate and candidate memory, including the hidden state h t-1 and input x t The linear transformation parameters of b i , b f , b o , b c It is the bias vector of each gate and candidate memory, adjusting the result of linear transformation.

[0109] Figure 2The LSTM neural network construction process in the present invention is shown. The specific parameters of the LSTM neural network are shown in Table 1:

[0110] Table 1 Specific parameters of LSTM neural network

[0111]

[0112] S4: defining a learning rate adjustment strategy for the LSTM neural network in step S3 to prevent the LSTM neural network from being unstable, specifically using a loss function optimization algorithm to optimize the learning rate and gradient clipping rate of the LSTM neural network in step S3 to obtain an optimized LSTM neural network, and then using an Adam (adaptive moment estimation) optimizer to train the optimized LSTM neural network to obtain a trained LSTM neural network;

[0113] By combining the LSTM neural network with real-time data adjustment strategy, the thrust distribution of multi-engine rockets in different flight phases is optimized, making the rocket's flight more stable and precise; the LSTM neural network can make thrust compensation adjustments based on real-time environmental data (such as pressure, wind speed, temperature, humidity, etc.), ensuring that even when a partial failure occurs in the rocket engine, the system can still reasonably distribute the thrust of the remaining engines to ensure flight safety and stability.

[0114] The parameters in the loss function optimization algorithm in step S4 are as follows:

[0115] The mean square error (MSE) is shown in formula (7):

[0116]

[0117] In formula (7), N is the total number of data, y true,t is the true value at time t, y pred,t is the predicted value at time t;

[0118] The angle change penalty term (change_penalty) is shown in formula (8):

[0119]

[0120] In formula (8), (y pred,t+1 -y pred,t ) is the change in the predicted value between time t and time t+1; (y true,t+1 -y true,t ) is the change in true value between time t and time t+1;

[0121] The angular velocity constraint (angle change rate) is shown in formula (9):

[0122]

[0123] If the rate of angle change exceeds a set maximum value, an angle rate change penalty (rate penalty) is applied, where rate penalty = max (0, angle change rate - max angle change), and in this embodiment, max angle change = 5°;

[0124] The total loss function (Total Loss) is shown in formula (10):

[0125] Total Loss = MSE + λ 1 change_penalty+λ 2 ·rate penalty (10)

[0126] In formula (10), λ1 and λ2 are the weight coefficients of change_penalty and rate_penalty, respectively, which control the influence of the penalty term.

[0127] In this embodiment, N=1000, t=1000s, λ1=0.4, λ2=0.6;

[0128] change_penalty is the penalty term for angle change, which encourages smooth posture changes, and rate_penalty is the penalty term for angle rate change, which ensures that the angle change rate does not exceed a certain threshold.

[0129] Adam itself has a mechanism for dynamically adjusting the learning rate, so there is no need to manually set the learning rate decay. However, in some cases, combining learning rate decay can further optimize the training process and improve the training effect.

[0130] The Adam optimization algorithm is a stochastic optimization method suitable for large-scale data and non-stationary targets. Its core idea is as follows:

[0131] Furthermore, the momentum update, second-order moment estimation, and parameter update in the Adam optimizer are shown in formulas (11), (12), and (13), respectively:

[0132] m t =β 1 m t-1 +(1-β 1 ) t (11)

[0133] In formula (11), m t is the momentum estimate, which indicates the cumulative information of the direction and magnitude of the gradient, mt-1 is the first moment estimate of the gradient (momentum), β 1 is the decay rate of the first-order moment, g t is the gradient at the current time. In this embodiment, β 1 is 0.9;

[0134]

[0135] In formula (12), v t is the second-order moment estimate, which represents the weighted sum of squares of the gradient magnitude and is used to adjust the update step size, v t-1 is the second-order moment estimate of the gradient, β 2 is the decay rate of the second-order moment, in this embodiment, β 2 is 0.999;

[0136]

[0137] In formula (13), θ t is the model parameter at the current time step, θ t+1 is the model parameter of the next time step, α is the global learning rate, which controls the step size of parameter update, ∈ is a smoothing term used to avoid zero division errors. In this embodiment, ∈ is 10 -8 ;

[0138] In this embodiment, the optimizer is the Adam optimizer, the learning rate is 0.001, the gradient clipping rate is 1.0, the LSTM output layer dimension is 100, and relu is used as the activation function;

[0139] Figure 4 The comparison between the validation loss rate and training loss rate of LSTM neural network and other models: fully connected feedforward neural network (MLP, Multi-Layer Perceptron), simple RNN (Recurrent Neural Network), 1D convolutional neural network (CNN) is shown. Figure 4It can be seen that the LSTM neural network can capture the long-term dependencies of time series data, and the training loss and validation loss are lower than other models. In this embodiment, the number of training repetitions is set to 200 times. After training with the LSTM neural network, the training loss rate (Training Loss) and the validation loss rate (Validation Loss) are effectively reduced; when the number of training (Epochs) is 100, the training loss rate and validation loss rate of the LSTM model are effectively lower than 0.1, and the training loss rate and validation loss rate of other models that do not use the LSTM model are between 0.1 and 0.4, indicating that the LSTM neural network model performs well throughout the training process. The stability of the LSTM neural network is significantly improved after the addition of the Adam optimizer, and it can effectively adapt to the validation set without obvious overfitting or underfitting, and also avoids the problem of output errors caused by model instability.

[0140] S5: Initialize the variables in step S1 and solve the dynamic variables of the launch vehicle through the dynamic equation. The dynamic variables of the launch vehicle include the total thrust F, acceleration The time-varying mass m(t), the derivative of the height h, dh, and gravity Attitude adjustment forces and launch vehicle thrust vector control torque;

[0141] The expression of the kinetic equation is shown in formula (14):

[0142]

[0143] In formula (14), m is the total mass of the rocket, m(t) is the sum of the mass of the rocket and the mass of the fuel at time t, and the thrust Air resistance gravity The total thrust F formed by the random disturbance determines the acceleration is the instantaneous combustion rate of the fuel; v e is the effective exhaust velocity, c e is the fuel consumption efficiency coefficient; is the rocket acceleration, F is the total thrust, v is the velocity, γ is the attitude angle, and h is the altitude;

[0144] Furthermore, the expressions of the attitude adjustment force and the launch vehicle thrust vector control torque are shown in formula (15):

[0145]

[0146] In formula (15), is the attitude adjustment force (single engine thrust changes with altitude), K pis the proportional gain coefficient, controlling the thrust compensation for the lateral wind speed, v wind is the wind speed, M tvc is the thrust vector control torque of the launch vehicle, Δθ is the attitude adjustment angle (i.e. ).

[0147] In this embodiment, the end time t is set 1 =1000s, attitude adjustment force The range is [100N, 10000N]. The rocket height increases with time. The attitude adjustment angle Δθ ranges from [85°, 92°]. The launch vehicle thrust vector control torque M tvc is [100N·m, 2000N·m], the thrust compensation for controlling the lateral wind speed is 0.7, and Δθ is the attitude adjustment angle.

[0148] S6: according to the environmental parameters initialized in step 2, using the environmental parameter expression to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, wind speed U;

[0149] The environmental parameter expression is shown in formula (16):

[0150]

[0151] In formula (16), ρ 0 , P 0 They represent the atmospheric density and atmospheric pressure at the surface respectively; k is the von Karman constant, which is 0.4, and u * is the friction velocity, U(z) is the average wind speed at height z; Γ is the vertical temperature gradient; T 0 is the sea level temperature; the temperature at the top of the troposphere (about 11 km) becomes constant; e(z) is the water vapor pressure at height z; e 0 is the water vapor pressure at sea level; M v is the molar mass of water vapor; R v is the gas constant of water vapor; T is the atmospheric temperature, T 0 is the initial atmospheric temperature, Γh is the product of the vertical temperature lapse rate and the variable, M v is the molar mass of water vapor, R v is the gas constant for water vapor.

[0152] In this embodiment, the environmental variables vary with the altitude, Γ = -6.5K / km, and the atmospheric density ρ ranges from [0.003kg / m 3 ,1.225kg / m 3], atmospheric pressure P range is [0.2kPa, 101.325kPa], temperature T range is [50.65K, 288.15k], water vapor pressure e range is [0.0000001kPa, 3kPa], wind speed U range is [10m / s, 200m / s];

[0153] Step 7: Input the dynamic variables and environmental variables calculated in step S5 and step S6 into the LSTM neural network trained in step S4 to simulate the attitude adjustment force and thrust torque of the launch vehicle, and obtain the flight simulation results of the launch vehicle to ensure that the attitude and thrust distribution of the rocket are reasonable under complex environmental changes.

[0154] Figure 5 The altitude change curve simulated by LSTM neural network is shown, which is more in line with physical characteristics. The initial stage is a slow climb, affected by air resistance and gravity; the height growth slows down significantly in the middle stage, simulating the transition area of ​​thrust reduction; the later stage is a rapid climb and tends to be stable, simulating the effect of reduced air resistance and increased thrust; Figure 6 The altitude change curve simulated without using LSTM neural network is shown. Only the changes in the initial stage and the middle stage are simulated. The middle stage is similar to the linear situation and cannot simulate the real flight process. Figure 7 It shows that after adding the LSTM neural network, the fuel consumption rate decreases, which can effectively save fuel; Figure 8 It shows that the speed change curve using the LSTM neural network takes into account dynamic environmental factors and realizes adaptive adjustment, which reflects better adaptability and anti-interference ability. Due to environmental disturbances, the attitude control force will change continuously. The attitude control force with the addition of the LSTM neural network is smoother and more stable, and the control range does not exceed ±5°. Fig. 9 , Fig.10 The attitude control angles include pitch, yaw and roll. Fig. 9 , Fig.10 By comparison, it can be found that the attitude control angle changes without the LSTM neural network are very chaotic, while the attitude control with the LSTM neural network is smoother and less prone to drastic fluctuations.

[0155] In the second aspect, a multi-form launch vehicle simulation system based on an LTSM neural network is provided, and the multi-form launch vehicle simulation method is applied. The simulation system comprises an initial parameter acquisition module, an LSTM neural network training module, an instantaneous parameter acquisition module and a rocket flight simulation module:

[0156] Initial parameter acquisition module: set the initial variables of the launch vehicle, including the initial mass m of the launch vehicle 0, Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Set the initialization environment parameters of the launch vehicle, the initialization environment parameters include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial temperature T 0 , air resistance F drag , turbulence disturbance I ` , initial atmospheric density ρ 0 ; Total mass of the rocket m = m 0 +m 1 +m Fuel ;

[0157] LSTM neural network training module: build an LSTM neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, an LSTM output layer dimension and an activation function are set; a loss function optimization algorithm is used to optimize the learning rate and the gradient clipping rate of the LSTM neural network to obtain an optimized LSTM neural network, and then an Adam optimizer is used to train the optimized LSTM neural network to obtain a trained LSTM neural network;

[0158] Instantaneous parameter acquisition module: According to the initialization variables, the dynamic variables of the launch vehicle are solved through the dynamic equation. The dynamic variables of the launch vehicle include total thrust F, acceleration The time-varying mass m(t), the derivative dh of the height h, and the gravity attitude adjustment force and launch vehicle thrust vector control torque; according to the initialized environmental parameters, using environmental parameter expressions to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, wind speed U;

[0159] Rocket flight simulation module: The dynamic variables and environmental variables are input into the trained LSTM neural network to simulate the attitude adjustment force and thrust torque of the launch vehicle to obtain the launch vehicle flight simulation result.

[0160] In a third aspect, an electronic device includes a memory and a processor:

[0161] Memory: used for storing a computer program for implementing the multi-modal launch vehicle simulation method based on the LTSM neural network;

[0162] Processor: used to implement the multi-modal launch vehicle simulation method based on LTSM neural network when executing the computer program.

[0163] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-morphological launch vehicle simulation method based on the LTSM neural network is implemented; the computer-readable storage medium includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store program codes.

[0164] The working principle of the present invention is:

[0165] This method constructs an LSTM neural network by initially setting basic parameters (i.e., the rocket initialization variables and initialization environment parameters in steps 1 and 2), optimizes the LSTM neural network parameters through a loss function optimization algorithm, and then iteratively trains the neural network based on the basic parameters of the input LSTM neural network and trains the neural network through an Adam optimizer. Finally, the parameters of the rocket during flight are simulated in the trained LSTM neural network, and the simulation results are obtained and output in the form of charts.

Claims

1. A multi-modal launch vehicle simulation method based on LTSM neural network, It is characterized in that The following steps are involved: S1: Set the initialization variables of the launch vehicle, including the initial mass m of the launch vehicle 0 , Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Total mass of the rocket m = m 0 +m 1 +m Fuel ; S2: Setting the initialization environment parameters of the launch vehicle, which include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial atmospheric temperature T 0 , air resistance F drag , Turbulence I ` , initial atmospheric density ρ 0 ; S3: Building an LSTM (Long Short-Term Memory) neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, a LSTM output layer dimension, and an activation function are set; S4: using a loss function optimization algorithm to optimize the learning rate and gradient clipping rate of the LSTM neural network in step S3 to obtain an optimized LSTM neural network, and then using an Adam (adaptive moment estimation) optimizer to train the optimized LSTM neural network to obtain a trained LSTM neural network; S5: Initialize the variables in step S1 and solve the dynamic variables of the launch vehicle through the dynamic equation. The dynamic variables of the launch vehicle include the total thrust F, acceleration The time-varying mass m(t), the derivative of the height h, dh, and gravity Attitude adjustment forces and launch vehicle thrust vector control torque; S6: according to the environmental parameters initialized in step S2, using environmental parameter expressions to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, and wind speed U; S7: The dynamic variables and environmental variables calculated in step S5 and step S6 are input into the LSTM neural network trained in step S4 to simulate the attitude adjustment force and thrust torque of the launch vehicle, and obtain the flight simulation result of the launch vehicle.

2. The simulation method according to claim 1, It is characterized in that The LSTM neural network construction process described in step S3 is as follows: S3.1: Obtain time series data from the initialization variables in step S1 and the initialization environment parameters in step S2, including time t (current time), speed v (current speed), initial atmospheric density ρ 0 , Single engine thrust Initial fuel mass m Fuel ; S3.2: Normalize the time series data in step S3.1, and normalize the data x norm As shown in formula (1): In formula (1), x is the unnormalized data, μ is the mean, and σ is the standard deviation; S3.3: Use the normalized data described in step S3.2 to build the LSTM neural network described in step S3, including an input gate, a forget gate, and an output gate.

3. The simulation method according to claim 1, It is characterized in that The parameters in the loss function optimization algorithm described in step S4 are as follows: The mean square error (MSE) is shown in formula (2): In formula (2), N is the total number of data, y true,t is the true value at time t, y pred,t is the predicted value at time t; The angle change penalty term (change_penalty) is shown in formula (3): In formula (3), (y pred,t+1 -y pred,t ) is the change in the predicted value between time t and time t+1; (y true,t+1 -y true,t ) is the change in true value between time t and time t+1; The angular velocity constraint (angle change rate) is shown in formula (4): If the angle change rate exceeds a set maximum value, an angle rate penalty is applied, where rate penalty = max(0, angle change rate - maxangle change) The total loss function (Total Loss) is shown in formula (5): Total Loss=MSE+λ 1 ·change_penalty+λ 2 ·rate penalty(5) In formula (5), λ1 and λ2 are the weight coefficients of change_penalty and rate_penalty, respectively, which control the influence of the penalty term.

4. The simulation method according to claim 1, It is characterized in that The momentum update, second-order moment estimation, and parameter update in the Adam optimizer in step S4 are shown in formulas (6), (7), and (8), respectively: m t =b 1 m t-1 +(1-β 1 )g t (6) In formula (6), m t is the momentum estimate, which indicates the cumulative information of the direction and magnitude of the gradient, m t-1 is the first moment estimate of the gradient (momentum), β 1 is the decay rate of the first-order moment, g t is the gradient at the current time; In formula (7), v t is the second-order moment estimate, which represents the weighted sum of squares of the gradient magnitude and is used to adjust the update step size, v t-1 is the second-order moment estimate of the gradient, β 2 is the decay rate of the second-order moment; In formula (8), θ t is the model parameter at the current time step, θ t+1 is the model parameter for the next time step, α is the global learning rate, which controls the step size of parameter update, and ∈ is a smoothing term used to avoid division by zero errors.

5. The simulation method according to claim 1, It is characterized in that The expression of the kinetic equation in step S5 is shown in formula (9): In formula (9), m is the total mass of the rocket, m(t) is the sum of the mass of the rocket and the mass of the fuel at time t, and the thrust Air resistance gravity The total thrust F formed by the random disturbance determines the acceleration is the instantaneous combustion rate of the fuel; v e is the effective exhaust velocity, c e is the fuel consumption efficiency coefficient; is the rocket acceleration, F is the total thrust, v is the velocity, γ is the attitude angle, and h is the altitude.

6. The simulation method according to claim 1, It is characterized in that The expressions of the attitude adjustment force and the launch vehicle thrust vector control torque in step S5 are shown in formula (10): In formula (10), is the posture adjustment force, K p is the proportional gain coefficient, controlling the thrust compensation for the lateral wind speed, v wind is the wind speed, M tvc is the thrust vector control torque of the launch vehicle, and Δθ is the attitude adjustment angle.

7. The simulation method according to claim 1, It is characterized in that The environmental parameter expression in step S6 is shown in formula (11): In formula (11), ρ 0 , P 0 They represent the atmospheric density and atmospheric pressure at the surface respectively; k is the von Karman constant, u * is the friction velocity, U(z) is the average wind speed at height z; Γ is the vertical temperature gradient; T 0 is the sea level temperature; the temperature at the top of the troposphere (about 11 km) becomes constant; e(z) is the water vapor pressure at height z; e 0 is the water vapor pressure at sea level; M v is the molar mass of water vapor; R v is the gas constant of water vapor; T is the atmospheric temperature, T 0 is the initial atmospheric temperature, Γh is the product of the vertical temperature lapse rate and the variable, M v is the molar mass of water vapor, R v is the gas constant for water vapor.

8. A multi-modal launch vehicle simulation system based on LTSM neural network, It is characterized in that The simulation method according to any one of claims 1 to 7 is applied, wherein the simulation system comprises an initial parameter acquisition module, an LSTM neural network training module, an instantaneous parameter acquisition module and a rocket flight simulation module: Initial parameter acquisition module: set the initial variables of the launch vehicle, including the initial mass m of the launch vehicle 0 , Single engine thrust Initial fuel burning rate α, altitude h, launch angle γ, configuration n, load m 1 , total mass of the rocket m, speed v, time t, initial fuel mass m Fuel ; Set the initialization environment parameters of the launch vehicle, the initialization environment parameters include turbulence parameter I, initial air pressure P 0 , wind speed v wind , initial temperature T 0 , air resistance F drag , turbulence disturbance I ` , initial atmospheric density ρ 0 ; Total mass of the rocket m = m 0 +m 1 +m Fuel ; LSTM neural network training module: build an LSTM neural network for rocket thrust compensation, in which an optimizer, a learning rate, a gradient clipping rate, an LSTM output layer dimension and an activation function are set; a loss function optimization algorithm is used to optimize the learning rate and the gradient clipping rate of the LSTM neural network to obtain an optimized LSTM neural network, and then an Adam optimizer is used to train the optimized LSTM neural network to obtain a trained LSTM neural network; Instantaneous parameter acquisition module: According to the initialization variables, the dynamic variables of the launch vehicle are solved through the dynamic equation. The dynamic variables of the launch vehicle include total thrust F, acceleration The time-varying mass m(t), the derivative dh of the height h, and the gravity attitude adjustment force and launch vehicle thrust vector control torque; according to the initialized environmental parameters, using environmental parameter expressions to solve the environmental variables during the rocket flight, the environmental variables include atmospheric density ρ, atmospheric pressure P, temperature T, water vapor pressure e, wind speed U; Rocket flight simulation module: The dynamic variables and environmental variables are input into the trained LSTM neural network to simulate the attitude adjustment force and thrust torque of the launch vehicle to obtain the launch vehicle flight simulation result.

9. An electronic device comprising a memory and a processor, Features: Memory: used for storing a computer program for implementing the multi-modal launch vehicle simulation method based on the LTSM neural network; Processor: used to implement the multi-modal launch vehicle simulation method based on LTSM neural network when executing the computer program.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-modal launch vehicle simulation method based on the LTSM neural network is implemented.

Images