Edge-cloud-intelligent-enhanced flexible hypersonic vehicle control method

Abstract

This invention belongs to the field of hypersonic vehicle control technology, and relates to an edge-cloud intelligent enhancement control method for hypersonic vehicles. The method includes constructing a longitudinal dynamic model of the hypersonic vehicle and transforming it into a control-oriented mathematical model to facilitate the design of the ESO (Elastomer Optimizer) and PID controller; designing an ESO with generalized elastic displacement observation to observe key system state parameters; designing a cascaded feedforward PID controller to control the attitude and velocity of the hypersonic vehicle; and designing an edge-cloud intelligent parameter tuning system to achieve online tuning of the PID control parameters. This method is an intelligent control method for hypersonic vehicles based on an elastic observer and has broad application prospects. Experimental results show that compared with traditional PID control, this method significantly improves tracking accuracy, response speed, and anti-interference capability, providing a new approach for the autonomous flight control of hypersonic vehicles.

Description

Technical Field

[0001] This invention belongs to the field of hypersonic vehicle control technology, and relates to an edge-cloud intelligent enhancement elastic hypersonic vehicle control method. Background Technology

[0002] During high-speed flight, hypersonic vehicles experience significant elastic vibrations and deformations. These deformations not only affect the vehicle's aerodynamic performance but also its control stability. Traditional control methods based on rigidity assumptions cannot accurately describe the impact of these elastic deformations on the control system, leading to reduced control precision and potentially causing structural damage. Therefore, it is essential to consider the vehicle's elastic state and employ more precise control strategies to ensure stable flight. By combining elastic expansion state observation with intelligent control algorithms, more precise and stable control of hypersonic vehicles can be achieved, which is of great significance for improving the performance and safety of hypersonic vehicles.

[0003] The patent "A Precise Anti-interference Tracking Controller for Flexible Hypersonic Vehicles" (CN102749851B) proposes a precise anti-interference tracking control strategy for flexible hypersonic vehicles. First, a longitudinal dynamic model of the flexible hypersonic vehicle is constructed. Second, a multi-source interference system model is built based on this longitudinal model: a control-oriented baseline model is established first, then the elastic mode is treated as interference and its interference model is constructed. Finally, the above models are combined, incorporating factors such as measurement deviation, parameter fluctuations, and external disturbances to form a multi-source interference system model. Then, a nonlinear interference observer is designed based on this multi-source interference system model to estimate the elastic mode interference. Finally, a precise anti-interference dynamic inverse composite tracking controller is constructed based on this nonlinear interference observer. This scheme has advantages such as strong anti-interference capability, fast tracking response, and high control accuracy, and is suitable for precise tracking control of altitude and speed of flexible hypersonic vehicles. Although this patent solves the anti-interference tracking control problem of flexible hypersonic vehicles to a certain extent, its design of estimating the influence of elasticity as interference is not accurate enough for the model, especially in complex and variable flight environments, thus reducing control performance. In addition, dynamic inverse control is highly dependent on model accuracy. However, during high-speed flight, the dynamic characteristics of hypersonic vehicles are affected by a variety of factors in extreme working environments, including aerodynamics, structural elasticity, and thermal effects in hypersonic flight environments. These factors make accurate modeling extremely difficult.

[0004] The patent "A Control Method for an Elastic Hypersonic Vehicle with Input Constraints and Input Disturbances" (CN114153144B) discloses a control method for an elastic hypersonic vehicle with input constraints and input disturbances. The composite control law consists of an anti-input saturation control law and an input disturbance separation compensation control law. The anti-input saturation control law includes nominal and auxiliary feedback control parts, used to control the elastic hypersonic vehicle under input constraints and parameter uncertainties. The input disturbance separation compensation control law is designed based on a disturbance state observer and is used to compensate for input disturbances experienced by the elastic hypersonic vehicle. Considering the constraints of control input, parameter uncertainty, and input disturbances, a composite control law including anti-input saturation and input disturbance separation compensation is designed for a hypersonic vehicle with elastic characteristics based on pipeline predictive control technology, thereby achieving stable flight and precise command tracking of the elastic hypersonic vehicle. Although this patent proposes a composite control law for elastic hypersonic vehicles and has achieved certain results in dealing with input constraints, parameter uncertainties, and input disturbances, the construction process of its composite control law is relatively complex, involving the coordinated design and parameter adjustment of multiple control links. In practical engineering applications, this often places high demands on the system's rapid response capability and also brings certain technical challenges to real-time dynamic adjustment.

[0005] The patent "Neural Network Control Method for Hypersonic Vehicles Based on Online Data Learning" (CN107479382B) proposes a neural network control method for hypersonic vehicles based on online data learning, which can solve the technical problem of poor stability in existing hypersonic vehicle control methods. However, the design process of this method is relatively cumbersome. The neural network control system needs to complete complex calculations in a very short time to ensure the real-time performance of control commands, which is not conducive to engineering implementation and affects the operational safety and control accuracy of the vehicle.

[0006] The paper "Elastic Hypersonic Vehicle Control Based on Sliding Mode Observer and Interference Observer [J]. Journal of Naval Aeronautical Engineering Academy, 2020, 35(03): 234-240" considers the difficult-to-measure angle of attack and track angle, as well as the influence of external interference such as elastic coupling and wind field. A sliding mode observer is designed to estimate the unknown angle of attack and track angle; an interference observer is designed to estimate the lumped interference including elastic coupling and external interference. Based on the estimates from the sliding mode observer and interference observer, the control surface deflection angle and fuel equivalent of the altitude subsystem and velocity subsystem are designed respectively to achieve effective tracking of the desired altitude and velocity signals. The chattering problem in the design of the sliding mode observer may affect the accuracy of the estimation of the unknown angle of attack and track angle, which may lead to deviations in the control surface deflection angle and fuel equivalent designed based on the estimates, making it impossible to accurately track the desired altitude and velocity signals. Meanwhile, when faced with complex and ever-changing elastic coupling and external disturbances, the interference observer may have certain errors in estimating the lumped disturbance, which will also reduce the performance of the control system to some extent and make it difficult to fully meet the control requirements of hypersonic vehicles in complex flight environments.

[0007] Existing technologies simply treat elastic effects as disturbances for estimation, which is insufficiently accurate in complex flight environments, reducing control performance. Furthermore, dynamic inverse control is overly dependent on model accuracy, and accurate modeling of hypersonic vehicles is extremely difficult. Constructing composite control laws is complex, involving the coordinated design and parameter adjustment of multiple control links, requiring high system responsiveness and posing technical challenges to real-time dynamic adjustments. In addition, the design steps of control methods are cumbersome, hindering engineering implementation and impacting vehicle operational safety and control accuracy. Moreover, disturbance observers exhibit estimation errors when facing complex disturbances, failing to fully meet control requirements in complex flight environments. Therefore, it is necessary to propose a novel intelligent control method for hypersonic vehicles based on elastic observers to better address these problems. Summary of the Invention

[0008] The purpose of this invention is to achieve stable tracking control of hypersonic vehicles based on an intelligent method of elastic state observation, providing an edge-cloud intelligent enhancement control method for hypersonic vehicles. By constructing and transforming a longitudinal dynamic model of hypersonic flight to facilitate subsequent observer and controller design, the elastic state of the vehicle is estimated in real time using the observer. The elastic state information estimated by the elastic expansion state observer is introduced into a cascaded PID controller. Based on different flight states and elastic changes, the PID control parameters are intelligently tuned using an edge-cloud deep reinforcement learning algorithm. This improves the adaptive capability and control accuracy of the control system, effectively suppressing the elastic state of the hypersonic vehicle and ensuring stable flight in complex environments.

[0009] The technical solution of the present invention is as follows:

[0010] An edge-cloud intelligently enhanced elastic hypersonic vehicle control method includes the following steps:

[0011] Step 1: Dynamic model and problem statement of an elastic hypersonic vehicle

[0012] Step 1.1: Construct a longitudinal dynamic model of the elastic hypersonic vehicle.

[0013] The longitudinal dynamic model of a hypersonic vehicle based on a free beam structure is represented as follows:

[0014]

[0015] In the formula, the superscript " " represents the first derivative, with a superscript " " denotes the second derivative; For flight speed, For the body mass, For thrust, As resistance, It is the acceleration due to gravity. For the track angle, For lift, For height, The pitch angle, For the angle of attack, The pitch angular velocity, For pitching moment, Let the pitch inertia be [value]. The rigid body state variables are [variables]. The elastomer is in a certain state. ; It is the damping ratio of the i-th order structure; and Let represent the i-th natural frequency and the elastic displacement in the generalized coordinate system, respectively. The force is the i-th order generalized force; the control inputs include the elevator deflection angle. Engine throttle opening Through force and torque , , , , It is applied to hypersonic vehicle models.

[0016]

[0017] in

[0018]

[0019] In the formula, For dynamic pressure, For reference area, For the thrust arm, For reference chord length, For elevators, , The lift coefficient, The drag coefficient, This is the pitch moment coefficient. For thrust coefficient, This is the fuel equivalence ratio term for the thrust coefficient. This is the angle-of-attack term of the thrust coefficient. This is the generalized elastic displacement term for the thrust coefficient. For the i-th order generalized force coefficient; The lift coupling coefficient is... The drag coupling coefficient is... This is the pitch moment coupling coefficient. The coupling coefficient of the fuel equivalence ratio term of the thrust coefficient. The coupling coefficient is the angle-of-attack term of the thrust coefficient. is the coupling coefficient of the i-th order generalized force coefficient.

[0020] Step 1.2: Transformation of the longitudinal dynamics model for a control-oriented elastic hypersonic vehicle

[0021] The complex hypersonic vehicle dynamics model described above is transformed into a form more suitable for control design. Addressing the issue that the rigid body and elastic body state variables in the model from step 1.1 are coupled and difficult to directly use for controller design, a state transformation matrix is ​​designed to recombine the state variables in the original model, separating out independent state variables directly related to rigid body motion and elastic vibration.

[0022] Through mathematical derivation, a linear transformation relationship between the original and new state variables was established, thereby achieving decoupling of the state space. In the decoupled state space, control inputs such as the combination of elevator yaw angle and engine throttle opening, and output variables such as the aircraft's attitude angle and velocity are defined. Finally, considering external disturbances and unmodeled parts of the system, a clear input-output model is established.

[0023] Define target height and tracking error The altitude and track cornering system is organized using the backstepping method, taking into account... and After sorting, we can obtain

[0024]

[0025] in

[0026]

[0027] In the formula, For the target track angle, , These are the track angle control parameters; For state coupling terms that include both known and unknown parts of the model, , , , and This refers to the model uncertainty caused by external disturbances to the channel and the unmodeled portion.

[0028] Step 2: Design of an intelligent observation and control system for an elastic hypersonic vehicle based on edge clouds

[0029] Step 2.1: Design the elastic expansion state observer

[0030] To achieve observation of disturbances and model uncertainties, the transformed model from step 1.2 is analyzed and an observer is designed:

[0031] First, as can be seen from the elastic subsystem of the transformed model, the solution of generalized displacement is highly dependent on the rigid body subsystem. Moreover, the solution process involves solving second-order differential equations, which will amplify the impact of uncertainty. Therefore, the observation gain term should be designed using data that can be directly measured by the sensor to avoid the uncertainty brought about by the generalized displacement in a series of calculations.

[0032] Next, a gain-compensated observer based on the transformed model is designed. To observe high-precision generalized elastic displacements, the generalized force in the observer's elastic subsystem equations includes state coupling terms and input terms. Angle-of-attack observations have a significant impact on the elastic system, being heavily influenced by the track angle and pitch rate, but less so by altitude. Furthermore, the track angle is highly measurable, resulting in minimal error that does not require observation. Therefore, the pitch rate and angle-of-attack observation errors are used as feedback to compensate for the observer's design. Simultaneously, an extended state is designed for the elastic system within the observer to observe unmodeled disturbances.

[0033] Finally, based on the above analysis, the elastic state expansion observer can be designed as follows:

[0034]

[0035] In the formula, the superscript " "Indicates an estimated value, This is the estimated angle of attack. This is an estimate of the pitch angular velocity. This is the estimate of the i-th order generalized elastic displacement. This represents the estimated value of the unknown disturbance in the elastic system equations. The observer gain coefficient, ~ This is the elastic disturbance compensation coefficient.

[0036] Step 2.2, Design of Cascade Feedforward PID Controller

[0037] A controller is designed using a state-extended observer. Based on the longitudinal model of the hypersonic vehicle transformed for control in step 1.2, a cascaded feedforward compensated PID attitude control law is designed. Simultaneously, using the observations of the generalized elastic displacement obtained from the observer designed in step 2.1, a model-based observation compensation term is designed to suppress the model uncertainty caused by elastic disturbances in the transformed surface model of step 1.2. This term, along with the observations, adjusts the elevator deflection and throttle opening, as follows:

[0038]

[0039] In the formula, , , The value is calculated by substituting the corresponding observer observations into the formula in step 1.2. For pitch angular velocity virtual control variable, Indicates the integration time; For the target pitch angle, For the target speed, , , These are pitch angle, pitch rate, and flight speed tracking error, respectively. , , , All of these are PID control parameters.

[0040] Step 2.3: Intelligent parameter tuning mechanism based on edge-cloud

[0041] At the intelligent tuning level, this invention employs a PID tuning method based on edge-cloud deep reinforcement learning.

[0042] First, let's outline the design concept. In deep reinforcement learning algorithms, one type, the Deep Deterministic Policy Gradient (DDPG) architecture, involves the Actor network adjusting actions based on the state and outputting PID parameters, while the Critic network evaluates the cumulative reward of each action. The state space includes the flight state and the current PID parameters, the action space contains the PID adjustments, and the reward function considers tracking error, control smoothness, and disturbance rejection. During training, the Actor and Critic interact with the environment: the Actor executes actions and observes the state and reward; the Critic updates the value function, and the Actor optimizes the policy. Through iterative training, the Actor learns to adjust its PID parameters to maximize the cumulative reward, achieving intelligent tuning.

[0043] In the edge-cloud deep reinforcement learning framework, this invention proposes that the cloud trains a cloud deep reinforcement learning strategy (C-DDPG) offline through a database. For the edge intelligent parameter tuning mechanism, this invention proposes an edge deep reinforcement learning strategy (E-DDPG). During system operation, the state of the edge PID controller is simultaneously input to both the cloud and the edge. The E-DDPG learns online through the edge digital twin control model, outputting edge tuning parameters, which are then fused with the C-DDPG parameters through weighted fusion to obtain the final PID control parameters. Based on the above ideas, the edge-cloud intelligent parameter tuning mechanism is designed as follows:

[0044] Step 2.3.1, Design of Terminal PID Controller

[0045] For the edge-cloud intelligent parameter tuning mechanism, the cascade feedforward PID control system designed in step 2.2 is reorganized into an edge PID controller:

[0046]

[0047] In the formula, , , , , t is the number of system runtime steps. This indicates the system step size to be set. These represent the proportional, integral, and derivative control parameter matrices, respectively.

[0048] Step 2.3.2, C-DDPG parameter tuning mechanism

[0049] Based on the streamlined terminal PID controller, design the C-DDPG architecture: Based on the PID controller, design the C-DDPG state input. as follows:

[0050]

[0051] In the formula, , It is an 18-dimensional natural number matrix.

[0052] The C-DDPG state input contains the aircraft's current state information, including tracking error, flight speed, pitch angle, pitch rate, and control inputs and outputs. This state information collectively constitutes the input to the reinforcement learning algorithm, used to evaluate the effectiveness of the current control strategy.

[0053] The C-DDPG action output is designed as follows:

[0054]

[0055] in, For cloud PID parameters, Indicates cloud actions, Indicates the cloud executor strategy. For cloud Actor networks, It is a 9-dimensional natural number matrix. The action output is the adjustment amount of the PID controller parameters, including the proportional coefficient, integral coefficient, and derivative coefficient. By continuously adjusting these parameters, the reinforcement learning algorithm can explore the optimal combination of PID control parameters under different flight conditions, thereby improving the adaptive capability and control accuracy of the control system. To ensure physical rationality, a tanh function combined with offset scaling is used for constraints.

[0056]

[0057] in, This is the original output of the Actor neural network. , These are control parameters The upper and lower limits are set. The final output is the cloud PID parameters tuned by C-DDPG. .

[0058] Next, we design the C-DDPG reward function. as follows:

[0059]

[0060] In the formula, The weights are for the reward function.

[0061] The cloud target Critic network is trained using gradient descent, and the loss function is... The design is as follows:

[0062]

[0063] Among them, the superscript " "This represents the parameter for the next state." Indicates the state In distribution The above indicates taking the expected value. Indicates cloud Critic network, This represents the cloud experience replay buffer. For cloud discount factor, This represents the cloud action value function.

[0064] The cloud target Actor network improves its policy based on the gradient signal provided by the Critic, and is designed as follows:

[0065]

[0066] in, For the objective function Regarding the gradient of the parameters of the cloud Actor network, For the state In distribution The above indicates taking the expected value. Output for the cloud Critic network Action exist Evaluate the partial derivative at the point. Actor Policy Network Regarding its parameters The gradient.

[0067] The gradient update design for the neural network is as follows:

[0068]

[0069] in, This indicates the weights for updating the cloud gradient.

[0070] Finally, the cloud digital twin control model uses the dynamic model designed in step 1.2, the observer designed in step 2.1, and the PID controller designed in step 2.2 to evaluate the training network of C-DDPG.

[0071] Step 2.3.3, E-DDPG parameter tuning mechanism

[0072] The following section designs the E-DDPG and parameter fusion tuning architecture. The E-DDPG architecture, like cloud reinforcement learning, employs a deep reinforcement learning strategy.

[0073] E-DDPG status input The design is as follows:

[0074]

[0075] The E-DDPG action output function is designed as follows:

[0076]

[0077] in, For edge PID parameters, Indicates the side action, Indicates the edge executor strategy. It is an edge Actor network.

[0078] The E-DDPG reward function is designed as follows:

[0079]

[0080] in, For the E-DDPG reward function, The input weight parameters are used. This reward function encourages rapid convergence in parameter tuning. Similarly, the target Actor and Critic networks of E-DDPG employ the same gradient update strategy as C-DDPG.

[0081] Loss function design for edge-target Critic networks

[0082]

[0083] Among them, the superscript " "This represents the parameter for the next state." Indicates the state In distribution The above indicates taking the expected value. Represents the edge Critic network. This represents the edge experience replay buffer. The edge discount factor, This represents the edge action value function.

[0084] The edge-target Actor network improves the policy based on the gradient signal provided by the Critic, and is designed as follows:

[0085]

[0086] in, For the objective function Regarding the gradient of the edge Actor network parameters, For the state In distribution The above indicates taking the expected value. For the output of the edge Critic network Action exist Evaluate the partial derivative at the point. Edge Actor Policy Network Regarding its parameters The gradient.

[0087] The gradient update design of the neural network is as follows:

[0088]

[0089] in, This indicates that the weights are updated using the edge gradient.

[0090] Simultaneously, the side-by-side digital twin control model uses the dynamic model designed in step 1.2, the observer designed in step 2.1, and the PID controller designed in step 2.2 to evaluate the training network of E-DDPG.

[0091] Step 2.3.4, Adaptive Weighting Mechanism

[0092] To combine the PID parameters tuned by C-DDPG and E-DDPG, a fusion weighted adaptive mechanism is designed as follows:

[0093]

[0094] In the formula, To integrate weights, It is the Sigmoid function. This is the sensitivity coefficient. The smaller the average error of the system attitude tracking, the greater the fusion weight; conversely, the greater the average error of the system attitude tracking, the greater the fusion weight.

[0095] Next, an adaptive bias is designed, and an online correction term is introduced.

[0096]

[0097] in, For online correction items, These are the parameters for the online correction term. The online correction term is driven by performance errors.

[0098] Finally, the parameter tuning function is:

[0099]

[0100] The beneficial effects of this invention are:

[0101] This invention achieves dynamic optimization and adaptive adjustment of control parameters for hypersonic vehicles through an edge-cloud intelligent parameter tuning mechanism. Specifically, this mechanism integrates reinforcement learning strategies from cloud-based C-DDPG and edge-based E-DDPG, significantly improving the control system's adaptability to complex flight environments through online learning and real-time feedback. The cloud-based C-DDPG utilizes global data and long-term experience to output basic PID parameters, while the edge-based E-DDPG outputs fine-tuning parameters based on local real-time data. The two mechanisms dynamically allocate control weights through a weight fusion mechanism, ensuring the physical rationality and real-time performance of parameter tuning. Furthermore, the introduction of an adaptive bias term further corrects performance errors during parameter tuning, enabling the control system to maintain high accuracy and robustness even in the face of sudden disturbances or model uncertainties. Experimental results show that this method significantly improves tracking accuracy, response speed, and anti-interference capability compared to traditional PID control, providing a new approach for the autonomous flight control of hypersonic vehicles. Attached Figure Description

[0102] Figure 1 This is a technical block diagram of an edge-cloud intelligent enhanced elastic hypersonic vehicle control method;

[0103] Figure 2 This is a detailed flowchart of the edge-cloud intelligent enhanced elastic hypersonic vehicle control method;

[0104] Figure 3 This is a schematic diagram of the edge-cloud PID intelligent parameter tuning mechanism;

[0105] Figure 4 This is a simulation curve of the speed tracking of a hypersonic vehicle;

[0106] Figure 5 This is a simulation curve of the altitude tracking of a hypersonic vehicle;

[0107] Figure 6 It is a graph showing the actual values ​​and observed values ​​of pitch rate and angle of attack of a hypersonic vehicle over time.

[0108] Figure 7 It is a time curve of the actual value and observed value of the generalized elastic displacement of a hypersonic vehicle;

[0109] Figure 8 This is a time curve of the observation error of the generalized elastic displacement of a hypersonic vehicle;

[0110] Figure 9 This is a time curve comparing the observation errors of the generalized elastic displacement of a hypersonic vehicle. Detailed Implementation

[0111] The embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0112] like Figure 1 and Figure 2 As shown, the edge-cloud intelligent enhanced elastic hypersonic vehicle control method of the present invention includes:

[0113] Step 1: Construct a longitudinal dynamic model of the elastic hypersonic vehicle and transform it into a control-oriented mathematical model to facilitate the design of ESO and PID controllers.

[0114] Step 2: Design an ESO that includes generalized elastic displacement observation to observe the key state parameters of the system.

[0115] Step 3: Design a cascaded feedforward PID controller to control the attitude and speed of the hypersonic aircraft.

[0116] Step 4: Design an edge-cloud intelligent parameter tuning system to realize online tuning of PID control parameters.

[0117] Specifically, the embodiments of the present invention are described as follows:

[0118] First, constructing a longitudinal dynamics model of the hypersonic vehicle is fundamental to the entire control method. This model needs to fully consider the vehicle's elastic characteristics, incorporating the impact of structural elasticity on flight dynamics into the model, thus obtaining a longitudinal dynamics model that includes elastic effects. This model not only accurately describes the vehicle's motion state but also facilitates the subsequent design of ESO and PID controllers. By making reasonable simplifications and assumptions, the model is transformed into a control-oriented mathematical model, providing an accurate model foundation for subsequent control design.

[0119] Next, an ESO (Employment Optimizer) incorporating generalized elastic displacement observations is designed. The ESO is a crucial component in the entire control system, addressing observation disturbances and model uncertainties. Based on the previously constructed longitudinal dynamic model, the state coupling terms and external disturbances in the model are analyzed, and appropriate observation gain terms are designed. Observers are designed using directly measurable data from sensors, such as the aircraft's speed, altitude, and pitch angle, to avoid uncertainties introduced by generalized displacements during calculation. For the elastic subsystem, considering that the generalized force includes state coupling terms and input terms, as well as the influence of angle-of-attack observations on elasticity, a third-order elastic state equation ESO is constructed to observe unknown, unmodeled disturbances. By rationally designing the ESO parameters, it can accurately estimate the aircraft's state and unknown disturbances, providing accurate feedback information for the subsequent PID controller.

[0120] Then, a cascaded PID controller is designed to control the attitude and velocity of the hypersonic vehicle. Based on the revised longitudinal model of the control-oriented hypersonic vehicle, a cascaded feedforward compensated PID attitude control law is designed. The cascaded control structure can improve the system's control accuracy and response speed, while feedforward compensation can compensate for known disturbances in advance, further improving the system's anti-interference capability. By observing the generalized elastic displacement, a model-based observation compensation term is designed to suppress disturbances caused by elasticity. According to the vehicle's flight state and the desired flight trajectory, the elevator deflection angle and throttle opening are adjusted so that the vehicle can accurately track the desired altitude and velocity signals. Appropriate PID control parameters are selected to ensure that the system has good dynamic and steady-state performance.

[0121] Finally, design an edge-cloud intelligent parameter tuning system, such as... Figure 2 This study aims to achieve online tuning of PID control parameters. An edge-cloud reinforcement learning-based PID tuning method is employed, with C-DDPG and E-DDPG architectures designed respectively. In the C-DDPG architecture, the state space is defined to include the current flight state information of the aircraft and the current values ​​of the PID control parameters. The action space is defined as the adjustment amount of the PID control parameters. The reward function comprehensively considers factors such as the aircraft's tracking error, the smoothness of the control input, and the suppression effect of elastic disturbances. Through continuous iterative training, the Actor network can learn how to adjust the PID control parameters in different states to obtain the maximum cumulative reward. The E-DDPG architecture also employs a deep reinforcement learning strategy, but focuses more on local parameter tuning. An adaptive weight fusion mechanism is designed, such as... Figure 3 The system automatically adjusts the fusion weights of cloud and edge PID parameters based on the magnitude of the average attitude tracking error, enabling it to better adapt to different flight conditions. Simultaneously, an adaptive bias is designed, introducing an online correction term. Performance error drives online parameter correction, further improving the accuracy and real-time performance of parameter tuning. Ultimately, the edge-cloud intelligent parameter tuning system achieves online tuning of PID control parameters, enhancing the control system's adaptability and control precision.

[0122] To verify the effectiveness of the intelligent control method for hypersonic vehicles based on elastic observers proposed in this invention, a comprehensive experimental study and comparative analysis were conducted.

[0123] The experiment selected a representative hypersonic vehicle as the research object and built a complete simulation experimental platform that included a vehicle dynamics model, an elastic observer, a cascaded PID controller, and an edge-cloud intelligent parameter tuning system. During the experiment, various flight conditions and external disturbance scenarios were set up to fully verify the performance of the control method under different conditions.

[0124] Design a longitudinal channel model of a hypersonic vehicle, and design the aerodynamic coefficients based on the Lisa model as follows:

[0125] Table 1 Aerodynamic Parameter Design Table

[0126]

[0127] The simulation time was set to 15 seconds, the simulation step size was 0.01 seconds, there was an internal system delay of 0.009249 seconds, the initial flight altitude was 2392.5 meters, and the flight speed was 2621.3 meters per second. The simulation was performed on the Matlab platform and the experimental results were obtained.

[0128] Figure 4 and Figure 5 The display shows that, under the set speed and altitude commands, with the help of the cascade feedforward PID controller and the edge-cloud intelligent parameter tuning system, both speed control and altitude control exhibit extremely high tracking accuracy and no obvious overshoot phenomenon, and can achieve stable tracking within 5 seconds. Figure 6 The results show that the time curves for both pitch rate and angle of attack exhibit good dynamic performance, with pitch rate control overshoot not exceeding 0.3 radians per second and angle of attack overshoot not exceeding 0.025 radians per second. This demonstrates that the design combining a cascaded feedforward PID controller with an elastic observer, and dynamically adjusting the PID control parameters through a weighted parameter tuning mechanism using C-DDPG and E-DDPG, effectively suppresses the impact of elastic disturbances on the aircraft's attitude. The aircraft can quickly and accurately track the desired velocity signal, and the velocity tracking error converges to a small range in a very short time and remains stable throughout the flight.

[0129] In addition, from Figure 6 It also shows that the designed elastic ESO observer achieves high-precision observation of pitch angular velocity and rapid observation of angle of attack within 1 second, with an observation error of no more than 0.05 radians. Figure 7 , Figure 8 , Figure 9 This further demonstrates that the elastic ESO can accurately estimate the aircraft's elastic state variables even in the presence of external airflow disturbances and model uncertainties, achieving precise observation of key state parameters, including the third-order generalized elastic displacement and its derivative terms. The observation error decays rapidly in the initial stage and converges to within the allowable error range within 5 seconds, verifying the robustness of the observer design. The elastic ESO effectively separates the state coupling terms from external disturbances through its third-order observation structure, ensuring a high degree of matching between observed and actual values, and providing a more reliable feedback basis for the control system.

[0130] Through detailed analysis and comparison of experimental data, the intelligent control method for hypersonic vehicles based on elastic observers proposed in this invention has been fully verified to have excellent performance in speed tracking, altitude tracking and attitude control. It can effectively improve the control accuracy, adaptability and robustness of hypersonic vehicles, and provide strong technical support for the safe and stable flight of hypersonic vehicles.

Claims

1. An edge-cloud-intelligent-augmented elastic hypersonic vehicle control method, characterized in that, Comprising the following steps: Step 1, the dynamic model of elastic hypersonic vehicle and problem statement; Step 1.1, the longitudinal dynamic model of elastic hypersonic vehicle is constructed; Step 1.2, the longitudinal dynamic model of elastic hypersonic vehicle for control is converted; In view of the problem that the rigid body state variables and the elastic body state variables are coupled in the model of step 1.1 and are difficult to be directly used for controller design, the state variables in the original model are recombined by designing a state transformation matrix, and independent state variables directly related to rigid body motion and elastic vibration are separated out; Through mathematical derivation, a linear transformation relationship between the original state variables and the new state variables is established, so that the state space is decoupled; in the decoupled state space, the control input is defined as the combination of elevator deflection angle and engine throttle opening, and the output variable is the attitude angle and velocity of the vehicle; finally, considering the external disturbance and the unmodeled part of the system, a clear input-output model is established; Step 2, intelligent observation and control system design of elastic hypersonic vehicle based on edge-cloud Step 2.1, design of elastic extended state observer In order to realize the observation of disturbance and model uncertainty, the converted model in step 1.2 is analyzed and an observer is designed: Firstly, the observation gain term is designed using the directly measurable data of the sensor; Then, the observer in the form of gain compensation based on the converted model is designed; The observer is compensated by using the pitch rate observation error and the angle of attack observation error as feedback; at the same time, the elastic system in the observer is designed to observe the unmodeled disturbance; Step 2.2, design of cascade feedforward PID controller According to the converted control-oriented hypersonic vehicle longitudinal model in step 1.2, a cascade feedforward compensation PID attitude control law is designed; at the same time, the observation value of the generalized elastic displacement is obtained by combining the observer designed in step 2.1, and a model-based observation compensation term is designed to suppress the model uncertainty caused by elastic disturbance in the converted model of step 1.2, so as to adjust the elevator deflection angle and the throttle opening; Step 2.3, edge-cloud intelligent parameter tuning mechanism At the intelligent tuning level, a PID tuning method based on edge-cloud deep reinforcement learning is adopted; In the edge-cloud deep reinforcement learning framework, it is proposed that the cloud trains the cloud deep reinforcement learning strategy C-DDPG offline through the database; for the edge intelligent parameter tuning mechanism, the edge deep reinforcement learning strategy E-DDPG is proposed; during system operation, the state of the edge PID controller is input to the cloud and the edge at the same time, E-DDPG learns online through the edge digital twin control model, and outputs the edge tuning parameters, which are fused with the C-DDPG parameters through weights to obtain the final PID control parameters.

2. The end-to-end cloud-intelligent-augmented flexible hypersonic vehicle control method according to claim 1, wherein, Step 1.1 is as follows: The longitudinal dynamic model of hypersonic vehicle based on free beam structure is represented as follows: wherein the superscript denotes the first derivative, and the superscript denotes the second derivative; is the flight velocity, is the mass of the body, is the thrust, is the drag, is the gravitational acceleration, is the track angle, is the lift, is the altitude, is the pitch angle, is the angle of attack, is the pitch rate, is the pitch moment, is the pitch inertia; The rigid body state variables are ; the elastic body state is ; is the i-th order structural damping ratio; and represent the i-th order natural frequency and the elastic displacement under the generalized coordinate, respectively; is the i-th order generalized force; the control input quantities include the elevator deflection angle , the engine throttle opening , the forces and torques , , , , acting on the hypersonic vehicle model; Wherein wherein is the dynamic pressure, is the reference area, is the thrust arm, is the reference chord length, is the elevator, , is the lift coefficient, is the drag coefficient, is the pitching moment coefficient, is the thrust coefficient, is the fuel equivalence ratio term of the thrust coefficient, is the angle of attack term of the thrust coefficient, is the generalized elastic displacement term of the thrust coefficient, is the i-th order generalized force coefficient; is the lift coupling coefficient, is the drag coupling coefficient, is the pitching moment coupling coefficient, is the coupling coefficient of the fuel equivalence ratio term of the thrust coefficient, is the coupling coefficient of the angle of attack term of the thrust coefficient, is the coupling coefficient of the i-th order generalized force coefficient.

3. The edge cloud intelligent augmented resilient hypersonic vehicle control method of claim 1, wherein, Step 1.2 is as follows: Define the target altitude and tracking error , the altitude and track angle subsystems are arranged using backstepping, and the and , the following is obtained Wherein wherein is the target track angle, , is the track angle control parameter; is the state coupling term containing the known and unknown parts of the model, , , , and are external disturbances to the channel and model uncertainties due to unmodeled parts.

4. The end-to-end cloud-intelligent-augmented flexible hypersonic vehicle control method of claim 1, wherein, The elastic state extended observer designed in step 2.1 is as follows: where the superscript represents an estimated value, is an estimated value of the angle of attack, is an estimated value of the pitch rate, is an estimated value of the i-th generalized elastic displacement, is an estimated value of the unknown disturbance in the elastic system equation, is an observer gain coefficient, ~ is an elastic disturbance compensation coefficient.

5. The edge cloud intelligent augmented resilient hypersonic vehicle control method of claim 1, wherein, The cascade feedforward PID controller in step 2.2 is as follows: wherein , , is the value calculated by substituting the corresponding observer observation value into the formula in step 1.2, is a virtual control amount of the pitch angle velocity, denotes an integration time; is a target pitch angle, is a target velocity, , , , respectively, are a pitch angle, a pitch angle velocity, and a flight velocity tracking error; , , , are all PID control parameters.

6. The end-to-end cloud-intelligent-augmented flexible hypersonic vehicle control method of claim 1, wherein Step 2.3 is as follows: Step 2.3.1, design of edge PID controller For the end-edge cloud intelligent parameter tuning mechanism, the cascade feedforward PID controller designed in step 2.2 is rearranged as an end PID controller: wherein , , , , , t is the number of system run time steps, is expressed as a set system step; respectively represent the proportional, integral, and derivative control parameter matrices; Step 2.3.2, C-DDPG parameter tuning mechanism According to the collated end PID controller, design C-DDPG architecture: according to the PID controller, design C-DDPG state input As follows: wherein , is an 18-dimensional natural number matrix; The C-DDPG state input contains the current state information of the aircraft, including tracking error, flight speed, pitch angle, pitch angle velocity, and control input and output; The C-DDPG action output function is designed as follows: where, is the cloud PID parameter, represents the cloud action, represents the cloud actor policy, is the cloud actor network, is a 9-dimensional natural number matrix; the action output is the adjustment amount of the PID controller parameters, including the proportional coefficient, the integral coefficient and the differential coefficient; in order to ensure physical rationality, the tanh function is combined with the offset scaling method for constraint: wherein, is the Actor neural network raw output; , are the upper and lower limits of the control parameters respectively set; the final output C-DDPG tuned cloud PID parameters ; Next, the C-DDPG reward function is designed As follows: In the formula, is a reward function weight; The cloud target Critic network is trained using a gradient descent method, with a loss function The design is as follows: where superscript denotes the parameter of the next state, denotes the state denotes the distribution denotes the expectation, denotes the cloud Critic network, denotes the cloud experience replay buffer, is the cloud discount factor, denotes the cloud action value function; The cloud target Actor network improves the strategy according to the gradient signal provided by the Critic, and is designed as follows: in, For the objective function Regarding the gradient of the parameters of the cloud Actor network, For the state In distribution The above indicates taking the expected value. Output for the cloud Critic network Action exist Evaluate the partial derivative at the point. Actor Policy Network Regarding its parameters The gradient; The neural network gradient update is designed as follows: wherein, denotes cloud gradient update weights; Finally, the cloud digital twin control model uses the dynamic model designed in step 1.2, the observer designed in step 2.1, and the PID controller designed in step 2.2 to evaluate the training network of C-DDPG; Step 2.3.3, E-DDPG parameter tuning mechanism E-DDPG state input The design is as follows: The E-DDPG action output function is designed as follows: wherein, is an edge PID parameter, represents an edge action, represents an edge effector policy, is an edge actor network; The E-DDPG reward function is designed as follows: wherein, is the E-DDPG reward function, is the input weight parameter; the target Actor and Critic networks of E-DDPG adopt the same gradient update strategy as C-DDPG; The loss function of the edge target Critic network is designed as follows where superscript represents the parameter of the next state, represents the state in the distribution represents taking the expectation, represents the edge Critic network, represents the edge experience replay buffer, is the edge discount factor, represents the edge action value function; The edge target Actor network improves the strategy according to the gradient signal provided by the Critic, and is designed as follows wherein, is the objective function is the gradient with respect to the edge Actor network parameters, is the state is the distribution denotes taking the expectation, is the edge Critic network output is the action evaluated at is the partial derivative, is the edge Actor policy network is the gradient with respect to its parameters ; The neural network gradient update is designed as follows wherein, denotes the edge gradient update weight; At the same time, the edge digital twin control model uses the dynamic model designed in step 1.2, the observer designed in step 2.1, and the PID controller designed in step 2.2 to evaluate the training network of E-DDPG; Step 2.3.4, fusion weight adaptive mechanism In order to combine the PID parameters tuned by C-DDPG and E-DDPG, the fusion weight adaptive mechanism is designed as follows: In the formula, is a fusion weight, is a Sigmoid function, is a sensitivity coefficient; the smaller the average error of system attitude tracking is, the greater the fusion weight is; the greater the average error of system attitude tracking is, the greater the fusion weight is; Then, the adaptive bias is designed, and an online correction term is introduced wherein, is an online correction term, are online correction term parameters, respectively; the online correction term is driven by the performance error; Finally, the parameter tuning function is: 。

Images