An experimental system for online autonomous monitoring and intelligent planning of space vehicles

Abstract

This application provides an experimental system for online autonomous monitoring and intelligent planning of space launch vehicles, comprising a user layer, a business layer, and a data access layer. The user layer provides a graphical user interface; the business layer receives and processes business requests from the user layer, performing data calculations, analysis, and mission planning based on these requests; and the data access layer manages local databases or data files, handling read / write and access requests for relevant data. The business layer includes a spacecraft simulation and fault data generation system, an online intelligent diagnostic system, and an online intelligent planning module. This system, through a multi-layered and modular design, achieves comprehensive monitoring, fault diagnosis, and intelligent planning of space launch vehicles.

Description

Technical Field

[0001] This application relates to the field of aerospace vehicle technology, and in particular to an experimental system for online autonomous monitoring and intelligent planning of a space launch vehicle. Background Technology

[0002] Space launch vehicles, as core equipment in the aerospace field, are widely used in critical missions such as satellite launches, deep space exploration, and manned spaceflight. The safety, stability, and efficiency of their flight process directly determine the success or failure of space missions. With the increasing complexity of space missions and the continuous rise in launch frequencies, the flight environment faced by launch vehicles is becoming increasingly dynamic and complex. Extreme space environments and the risk of sudden failures place higher demands on the online monitoring, fault diagnosis, and intelligent planning capabilities of launch vehicles. Developing efficient, reliable, and flexible online autonomous monitoring and intelligent planning experimental systems has become an important research direction and urgent need in the current aerospace field.

[0003] Currently, significant progress has been made in the monitoring and planning technologies for space launch vehicles. Some systems integrate cutting-edge technologies such as big data, artificial intelligence, and digital twins, attempting to achieve intelligent scheduling, fault early warning, and process assurance for launch missions. However, the following problems still exist: First, fault diagnosis relies on single sensor data or shallow models, making it difficult to capture complex nonlinear fault characteristics and prone to false alarms and missed alarms. Second, trajectory and mission reconstruction mostly adopt offline planning or traditional numerical optimization methods, resulting in poor real-time performance and insufficient consideration of trajectory deviations caused by fault uncertainties. Third, control law reconstruction lacks deep coordination with mission and launch trajectory reconstruction, and has insufficient adaptability to power redundancy after a fault. Fourth, the system is mostly a discrete architecture, with high data interaction latency and low integration of intelligent algorithms with flight control closed loop, making it difficult to achieve full-process autonomous closed loop. Summary of the Invention

[0004] In view of the above problems, this application provides an experimental system for online autonomous monitoring and intelligent planning of space launch vehicles, which aims to achieve comprehensive monitoring, fault diagnosis and intelligent planning of space launch vehicles, so as to overcome the above problems or at least partially solve the above problems.

[0005] This application provides an online autonomous monitoring and intelligent planning experimental system for space launch vehicles. The system includes a user layer, a service layer, and a data access layer, wherein: The user layer is used to provide users with a graphical user interface. The business layer is used to receive and process business requests from the user layer, and to perform data calculations, analysis, and task planning based on the business requests. The data access layer is used to manage local databases or data files and handle the reading, writing, and access requirements of related data.

[0006] Furthermore, the system also includes hardware components, wherein the hardware components include: The onboard computer is used to perform various flight control, mission management, and fault diagnosis tasks. Equivalent simulator, used to simulate spacecraft flight status and fault scenarios; Functional boards are used to implement data acquisition, control execution, and signal processing; Sensors and actuators collect flight data from the launch vehicle through different sensors and control the attitude adjustment of the spacecraft through actuators.

[0007] Furthermore, the feature is that the business layer includes: The spacecraft simulation and fault data generation system is used to simulate the flight process of launch vehicles and can simulate typical faults that rockets may encounter during flight. The online intelligent diagnostic system is used to monitor the onboard status and perform online fault diagnosis based on key data output by the system generated from spacecraft simulation and fault data. The online intelligent planning system is used to re-divide the launch trajectory mission area and re-plan the launch trajectory and control law when a launch vehicle malfunctions, and adjust the launch vehicle's flight attitude and orbit.

[0008] Furthermore, the spacecraft simulation and fault data generation system includes: The fully digital simulation module for the flight process is used to simulate the six-degree-of-freedom ballistic motion, dynamic characteristics, aerodynamic characteristics, and external environment of a spacecraft, and obtain the flight status data of the spacecraft. The sensor data generation module is used to generate key sensor simulation data based on the spacecraft's flight status data and key sensor models; The typical fault modeling module is used to model potential fault conditions based on simulation data from key sensors.

[0009] Furthermore, the online intelligent diagnostic system includes: The autonomous health assessment module is used to preprocess the acquired flight status data and sensor simulation data of the spacecraft, and to monitor the status based on the preprocessed standardized data, and to issue alarms for any abnormal states. An autonomous fault identification module is used to analyze and identify received abnormal state alarm information to obtain fault classification results; the fault classification results include fault type, location and severity. The fault handling decision module determines the fault index vector based on the fault classification results, and maps the fault index vector into fault handling instructions through a decision mapping function.

[0010] Furthermore, the preprocessing of the acquired sensor data includes: For raw sensor data with uneven distribution, the Lagrange polynomial interpolation algorithm is used to fill in the missing parameters; An improved wavelet threshold function is constructed by introducing exponential functions with different bases. The incomplete rocket sensor sample data is denoised based on the improved wavelet threshold function.

[0011] Furthermore, the improved threshold function is: in, As a regulating factor, T For the threshold, To improve the threshold function after processing the first... The first in scale Wavelet coefficients, To improve the threshold function before processing the first... The first in scale Wavelet coefficients.

[0012] Furthermore, the decision mapping function is: Among them, each parameter matrix With threshold All data were obtained through offline data analysis and online adaptive correction. To provide comprehensive fault indicators, This is the baseline value vector for each indicator.

[0013] Furthermore, the online intelligent planning module includes: The launch trajectory mission reconfiguration module is used to determine the reachability domain of the launch vehicle in an abnormal state when the launch vehicle malfunctions, and to repartition the mission based on the reachability domain data. The launch trajectory reconstruction module is used to replan the launch trajectory based on the mission area and mission partitioning, and reconstruct the optimal launch trajectory for the launch vehicle to complete the original intended orbital mission in the event of a sudden failure. The control reconfiguration module is used to reconfigure the flight trajectory control law and attitude control law, thereby controlling the launch vehicle's flight attitude and trajectory in the event of a sudden failure.

[0014] The specific beneficial effects are as follows: First, the system architecture is reliable and highly adaptable. It adopts a three-layer architecture (user layer, business layer, and data access layer), and achieves collaboration between the layers through standardized interfaces. It can operate stably in complex aerospace mission scenarios. At the same time, it has the advantages of high stability, convenient deployment, low maintenance difficulty, and the layered modular design provides a good foundation for subsequent hot-swappable modules, algorithm integration and function customization, adapting to the needs of technological evolution and function expansion. Second, the spacecraft simulation and fault data generation system of this system can realize the full digital simulation of the launch vehicle flight process, accurately simulate the characteristics of the rocket body, the external environment and the navigation and guidance control system, and at the same time model the key sensors and measurement equipment of the entire rocket, simulate various typical faults and output key data, providing comprehensive and accurate data input support for fault diagnosis, mission planning and other modules. Third, the online intelligent diagnostic module of this system adopts an intelligent fault diagnosis scheme that combines data and models, which can realize fault diagnosis at the entire rocket level, accurately identify faults in the control system, structural system and power system, and issue timely processing instructions; the top-level flight control software can perform diverse fault handling operations based on the diagnostic results, effectively improving the system's fault response capability and ensuring operational safety. Fourth, the online intelligent planning module of this system has the functions of mission reconfiguration, launch trajectory replanning and control reconfiguration. Based on the actual flight status and graded fault diagnosis results, it can judge the orbit reachability and adjust the flight launch trajectory in real time. At the same time, it realizes fault-tolerant control under fault conditions through power redundancy online reconfiguration technology, which significantly improves the system's adaptability, reliability and safety under abnormal conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a general framework diagram of the experimental system for online autonomous monitoring and intelligent planning of space launch vehicles provided in the embodiments of this application; Figure 2 This is a logical structure diagram of the experimental system for online autonomous monitoring and intelligent planning of space launch vehicles provided in the embodiments of this application; Figure 3 This is the data composition of the online autonomous monitoring and intelligent planning experimental system for space launch vehicles provided in the embodiments of this application; Figure 4 This is a hardware architecture diagram of the experimental system for online autonomous monitoring and intelligent planning of space launch vehicles provided in the embodiments of this application; Figure 5 This is a framework diagram of the spacecraft simulation and fault data generation system provided in the embodiments of this application; Figure 6 This is a framework diagram of the online intelligent diagnostic system provided in the embodiments of this application; Figure 7 This is the original signal provided in the embodiments of this application without threshold processing; Figure 8 This is the signal processed by the new threshold function proposed in this invention, as provided in the embodiments of this application; Figure 9 This is a flowchart of the Bayesian optimization process provided in the embodiments of this application; Figure 10 This is the interface for fault diagnosis results of the inertial navigation system provided in the embodiments of this application; Figure 11 This is the power system fault diagnosis result interface provided in the embodiments of this application; Figure 12 This is a framework diagram of the online intelligent planning system provided in the embodiments of this application; Detailed Implementation Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0017] Reference Figure 1 , Figure 1 This application provides an overall framework diagram of an experimental system for online autonomous monitoring and intelligent planning of a space launch vehicle. Its overall architecture includes a data layer, a model layer, an algorithm layer, an application layer, and a security layer. Through interaction between these different layers, comprehensive monitoring, fault diagnosis, and intelligent planning of the space launch vehicle are achieved. Wherein: The data layer is used to collect and store basic data in real time and preprocess it; the basic data includes flight status, sensor data, and fault diagnosis information. First, real-time flight status data (such as position, velocity, attitude, temperature, and pressure) of the spacecraft is collected by key distributed sensors throughout the rocket (e.g., IMU, GPS, optical sensors). Preliminary preprocessing (e.g., noise reduction, filtering) is then performed to ensure data quality. Subsequently, the collected data is stored in a distributed database, allowing users or the system to efficiently query and access historical data in real time. Then, the data layer transmits the collected data to the model layer, algorithm layer, and application layer via standardized communication protocols. A reasonable data flow mechanism ensures seamless data integration within the system.

[0018] The model layer is used by the algorithm layer to provide accurate flight simulation and fault prediction models; Specifically, the model layer serves as a bridge between the data layer and the algorithm layer. Based on real-time data acquired from the data layer, it first simulates the spacecraft's flight process using mathematical models, evaluating key parameters such as flight trajectory, fuel consumption, and velocity changes under different control inputs. The simulation results can be used for spacecraft flight analysis and trajectory optimization. Subsequently, based on known system failure modes (such as propulsion system failure, structural system failure, navigation and control system failure, electrical system failure, separation system failure, and environmental suitability failure), failure models are established for simulation, modeling the spacecraft's operational state under failure conditions. These models are used for system failure detection and prediction, helping the system make rapid decisions when failures occur. Finally, based on mission requirements and system status, the tools provided by the model layer are used for mission planning operations such as mission reconfiguration, launch trajectory replanning, and control reconfiguration. By simulating the results of different planning strategies, the model layer selects the optimal execution path for the system.

[0019] The algorithm layer is responsible for the computation of intelligent decision-making and fault handling; Specifically, the algorithm layer can monitor the health status of spacecraft and detect faults based on real-time data and advanced diagnostic algorithms (such as machine learning and neural networks); it can also use intelligent planning algorithms to plan and optimize mission paths. When a fault occurs, the algorithm layer dynamically adjusts the mission path (or degrades the orbit) based on real-time diagnostic results to ensure that the rocket can continue to perform the mission in the changed environment; it also provides control algorithms to replan the flight control strategy when the spacecraft malfunctions to ensure that the spacecraft can still fly stably.

[0020] The application layer provides users with a convenient user interface and real-time feedback; The application layer is primarily responsible for the interaction between the user and the system, providing a graphical interface to display information such as the spacecraft's real-time status, mission progress, fault diagnosis, and optimization suggestions. Users can view the rocket's flight information and mission execution status at any time. In addition, users can configure space missions through the interface and adjust mission plans as needed. Through the intuitive operation interface, users can quickly modify flight trajectories, replan missions, or activate emergency modes. When the system diagnoses a fault, the application layer will promptly issue an alarm to the user and provide intelligent decision support to help the user decide on the next steps (such as initiating fault handling or adjusting the mission).

[0021] The security layer is used to ensure the confidentiality and reliability of the system.

[0022] The security layer not only needs to protect the privacy of user data and flight data, but also needs to ensure the security and consistency of its control commands during rocket flight to prevent the system from being maliciously tampered with or leaked data.

[0023] The close cooperation between these layers enables the system to respond flexibly to dynamic and complex flight environments and provide optimal solutions.

[0024] Furthermore, based on the above-mentioned hierarchical design, this system achieves efficient operation, good scalability, and easy maintenance in a network-free environment by rationally dividing functional modules and hierarchical design. Figure 2 This is the logical architecture diagram of the system, refer to Figure 2 This system comprises a user layer, a business layer, and a data access layer; among which: The user layer is used to provide users with a graphical user interface. Specifically, users can interact with the system through a local graphical user interface, thereby easily accessing the system's functional modules and obtaining the information they need; The business layer is used to receive and process business requests from the user layer, and to perform data calculations, analysis, and task planning based on the business requests. The data access layer is primarily responsible for managing local databases or data files, and handling read, write, and access requests for related data.

[0025] The data access layer interacts with the business layer to provide it with the necessary data support and ensures high availability, consistency, and security of the data. Users can initiate data query requests and processing to the data access layer through the user layer, accessing the business layer. The data access layer returns the required content based on the query conditions, ensuring the system can quickly obtain relevant information during real-time task processing. The data stored and managed in the data access layer includes real-time flight data, mission data, fault diagnosis data, control command data, and historical data, such as... Figure 3 As shown.

[0026] Furthermore, the hardware framework of this system is as follows: Figure 4 As shown, its hardware components include the launch vehicle's onboard computer, equivalent simulator, function boards, sensors, and actuators. Among them: The onboard computer is the core computing unit of a spacecraft, responsible for performing various tasks such as flight control, mission management, and fault diagnosis. Onboard computers typically consist of multiple processing units to support efficient multitasking and processing requirements. They connect to other subsystems (such as sensors, control systems, and actuators) via a bus to exchange data and transmit commands.

[0027] An equivalent simulator is a hardware device in the system used to simulate the flight state and fault scenarios of a spacecraft. It is mainly used to simulate various flight environments, fault modes, and spacecraft behaviors in actual flight, providing reliable data input for system decision-making, fault diagnosis, and mission planning.

[0028] Functional boards are core components supporting various hardware functions of spacecraft. They mainly include hardware modules for data acquisition, control execution, and signal processing. Different functional boards are responsible for different subsystems, such as propulsion system control, sensor data acquisition, and attitude control. Among them, data acquisition boards are mainly responsible for acquiring data from sensors and transmitting the data to the onboard computer through data transmission interfaces. Typical examples include temperature sensor boards, pressure sensor boards, and accelerometer boards. Control execution boards, on the other hand, execute corresponding tasks by driving control elements such as servos and thrusters according to the control commands issued by the system.

[0029] Sensors and actuators are the sensing and execution components of a system, directly related to the spacecraft's operational status and mission execution. Sensors collect flight data, and actuators adjust the spacecraft's state according to commands. Key sensors include: an inertial measurement unit (IMU) for measuring motion parameters such as attitude, acceleration, and angular velocity; positioning sensors for acquiring precise spacecraft position data; temperature and pressure sensors for monitoring the environmental conditions of various spacecraft components; and (accelerometer) sensors for real-time acquisition of the rocket's flight status. Actuators, primarily based on the rocket's attitude and flight trajectory, are divided into servo motors and the propulsion system. Servos control the spacecraft's attitude adjustment, while the propulsion system provides power control for the spacecraft.

[0030] Furthermore, the business layer includes the following subsystems: (1) Spacecraft simulation and fault data generation system The spacecraft simulation and fault data generation system is used to simulate the flight process of launch vehicles and can simulate typical faults that the rocket may encounter during flight; such as... Figure 5 As shown, it includes the following modules: ①Full digital simulation module for flight process The fully digital simulation module for the flight process is used to accurately model and simulate the six-degree-of-freedom launch trajectory motion, dynamic characteristics, aerodynamic characteristics, and external environmental influences of the launch vehicle. Through high-precision simulation of the entire rocket flight, it provides the system with key flight state data such as attitude, velocity, and acceleration.

[0031] The full digital simulation of the launch vehicle's flight process mainly includes: First, selecting a coordinate system that can describe the forces and motion of the launch vehicle and perform a comprehensive and detailed dynamic and kinematic analysis; second, comprehensively considering the effects of multi-source perturbations, analyzing the forces acting on the launch vehicle, and constructing high-precision kinematic and dynamic equations for the launch vehicle; finally, designing different types of solution algorithms according to the requirements of multiple mission scenarios to efficiently and quickly solve the constructed kinematic and dynamic equations, and defining and transforming the coordinate system for the launch vehicle's motion and force analysis.

[0032] Specifically, a launch vehicle flying within Earth's gravitational field is subject to the effects of engine thrust and control forces, aerodynamics, and torque, as well as Earth's gravity. Earth's gravity is the primary force acting on the launch vehicle, significantly influencing its motion. The accuracy of the calculated Earth gravity value directly impacts the precision of the rocket's full digital simulation. Considering the launch vehicle's proximity to Earth and its relatively great distance from the Sun and other celestial bodies, the influence of the Sun and other celestial bodies on the rocket's motion can be ignored. Apart from the influence of Earth's rotation around its axis, other Earth motions have a minor impact on the launch vehicle's flight and are also negligible.

[0033] In launch orbital mechanics, the Earth is approximated as a sphere with a uniform mass distribution. According to Earth's gravity, the work done by the Earth's gravitational pull on any point in outer space depends only on that point's location and is independent of its path. Therefore, when the Earth's gravitational field is a positive force field, the work done by the Earth on any point outside a uniform sphere... The gravitational potential function can be expressed as: In the formula, Any point outside Earth quality It is the distance from a point to the center of the Earth. It is the gravitational constant; For any point mass per unit mass outside the Earth, its gravitational potential coefficient is: From the above equation, it can be seen that when the Earth is a homogeneous sphere, the gravitational potential at any point outside it is equal to the gravitational potential at points outside it if all its mass were concentrated at the center. Based on the property that the partial derivative of the gravitational potential with respect to any direction is equal to the potential function of the gravitational force in that direction, the magnitude of the gravitational acceleration per unit mass of a point mass when the Earth is a sphere can be expressed as: Its direction is from point P towards the Earth's center.

[0034] When the Earth's average radius And the average gravitational acceleration of the Earth is At that time, the magnitude of the gravitational acceleration applied to the rocket can be expressed as: Gravitational acceleration vector Decomposing the coordinates along each axis of the inertial coordinate system yields: (5) In the formula, These represent the gravitational accelerations in three directions. The distances are in the three directions in the inertial coordinate system.

[0035] Taking Earth's rotation into account, in addition to being affected by gravity, the rocket is also subject to Earth's gravity. In addition, it is also subject to the centrifugal inertial force generated by the Earth's rotation. Therefore, the gravitational force acting on the rocket G It is the vector sum of gravity and centrifugal inertial force, that is The Earth model uses a normal ellipsoid (IAG-75 normal Earth ellipsoid), and the following forces are all calculated in the launch coordinate system: the direction of gravity is consistent with the direction of the plumb line, i.e., the normal to the ground at the object's location. They are collinear and in opposite directions.

[0036] Calculations show that centrifugal inertial force Compared to gravity The magnitude is much smaller, therefore, gravity is usually considered... Treating it as gravity, that is, simplifying it to ; The projection of Earth's gravity can be represented as: in, The gravitational accelerations at... The geocentric vector has two components: one in the direction of rotation and the other in the direction of Earth's rotation. The unit vector of the geocentric vector is: (7) in, In the formula, These represent the distance from the Earth's center to the launch point, the launch point's elevation, the launch point's geodetic latitude, the launch point's geocentric latitude, and the geodetic azimuth of the launch aiming direction. These are the Earth's oblateness and the Earth's equatorial radius, respectively.

[0037] The unit vector of the Earth's rotational angular velocity vector is: (8) Based on the Earth's rotational angular velocity vector, the components of Earth's gravitational acceleration in the launch coordinate system can be obtained as follows: (9) Based on this, and according to the above analysis of the gravitational forces acting on the launch vehicle, simulation data of the Earth's gravitational field were obtained, as shown in the following figures. Figure 2 As shown.

[0038] Aerodynamics refers to the forces exerted by air on an object moving within it. When a compressible viscous airflow passes over the surfaces of various components of a launch vehicle, a pressure difference occurs due to the asymmetrical pressure distribution across the surface. Furthermore, the viscous friction of the atmosphere on the launch vehicle surface generates a corresponding viscous frictional force. These two forces combine to form the aerodynamic forces acting on the rocket.

[0039] The fundamental atmospheric properties affecting launch vehicle flight are primarily density, pressure, and temperature. These parameters depend mainly on the rocket's altitude, latitude, season, and day / night cycle. Standard atmospheric models approximate the average vertical distribution of atmospheric parameters such as density, pressure, and temperature using simplified methods based on measured data. This system adopts the USSA76 atmospheric model, commonly used in aerospace engineering. This model uses latitude as a baseline. The sea level of the region is used as a benchmark, and the statistical average of the actual atmospheric parameters throughout the year is used as the standard atmospheric parameters.

[0040] The aerodynamic force is decomposed into three components along the velocity coordinate system. The magnitude of the aerodynamic force is related to the dynamic pressure coefficient of the airflow. The characteristic area (also known as the reference area) of a rocket. Proportional, that is: In the formula, X , Y , Z These are drag, lift, and lateral force, respectively. These are dimensionless proportionality coefficients, referred to as drag coefficient, lift coefficient, and lateral force coefficient (collectively known as aerodynamic coefficients). air density, For the rocket's flight speed, For reference area, the area of ​​the arrow wing or the maximum cross-sectional area of ​​the arrow shaft is usually taken; Given Analyzing the aerodynamic forces acting on a rocket during flight can be simplified to studying the factors of these aerodynamic forces. To facilitate the analysis of the rocket's rotational motion, the total aerodynamic torque is... Along the projectile coordinate system It can be decomposed into three components, which are called rolling torques. (and (When the positive directions of the axes are aligned, the yaw moment is defined as positive). (and When the positive directions of the axes are aligned, it is defined as positive (and pitching moment). (and When the positive directions of the axes are aligned, it is defined as positive. The expression for aerodynamic torque is: in, These are dimensionless proportionality coefficients, referred to as roll coefficient, yaw moment coefficient, and pitch moment coefficient (collectively referred to as aerodynamic moment coefficients). For characteristic length, the missile body length is usually selected as the characteristic length in engineering applications, but the wingspan or mean aerodynamic chord length can also be used as the characteristic length.

[0041] When a rocket rotates around its axes in the projectile coordinate system, it generates corresponding damping moments. These damping moments are typically expressed as derivatives of their coefficients. The formulas for calculating their three components in the projectile coordinate system are as follows: in, These are respectively called the derivatives of the rolling damping moment coefficient, the yaw damping moment coefficient, and the pitch damping moment coefficient, and their values ​​can be determined by experiments. Let be the dimensionless pitch rate, dimensionless yaw rate, and dimensionless roll rate, defined as follows: in, These are roll rate, yaw rate, and pitch rate, respectively.

[0042] This embodiment uses a liquid rocket engine as an example to analyze the factors affecting thrust. For liquid rocket engines, the main factors affecting thrust are propellant flow rate per second, propellant mixture ratio, and specific thrust.

[0043] In summary, the formula for calculating the actual thrust of a rocket engine during flight can be expressed as: in, The actual and rated values ​​of the thruster's weight per second consumption; This refers to the actual and rated values ​​of oxidant consumption per second by weight; This refers to the actual and rated values ​​of the fuel weight per second consumption. The amount by which the deviation in propellant weight per second consumption affects the thrust; This is the amount by which the propellant mixture ratio deviation corrects for the thrust. This is the amount by which the thrust deviation is corrected for the thrust. Engine efficiency is a function of the propellant mixture ratio. This is the rated mixing ratio of the propellant, a constant coefficient that can be found in the engine test datasheet.

[0044] A rocket flies according to certain relationships under the control of a control system. To change the rocket's flight state, the magnitude and direction of the external forces acting on it must be altered. The main forces acting on a rocket are aerodynamic forces. ,thrust and gravity Due to gravity Since it always points towards the Earth's center, and its size and direction cannot be changed arbitrarily, controlling the rocket's flight can only rely on altering its aerodynamics. and thrust Their combined force becomes the controlling force. The formula for calculating control force is: Changes in velocity are typically achieved through thrust control, i.e., through normal control force. To change the direction of velocity, i.e., the rocket's flight direction, the calculation formula is: in, These are the normal component of thrust, lift, and lateral force, respectively. Finally, by adapting efficient and high-precision numerical analysis methods to different scenario mission requirements, a fully digital simulation of the launch vehicle flight process is achieved; the numerical analysis methods include Euler method, second-order Runge-Kutta method, fourth-order Runge-Kutta method, etc.

[0045] ② Sensor data generation module The sensor data generation module is used to generate sensor data, inertial measurement data and onboard BIT data during the launch and flight phases by modeling key distributed sensors and key measurement devices throughout the rocket. Specifically, during rocket flight, sensor data from the propulsion system is crucial for real-time monitoring and fault diagnosis. To effectively generate sensor data for propulsion system fault detection, the entire rocket's propulsion system exhibits clear modular and sequential characteristics: the first stage utilizes parallel engine technology to boost thrust to the hundreds of tons, providing strong power for liftoff and initial climb; the second stage employs a combination of four engines, responsible for precise attitude and orbit control during the second-stage flight phase; the third-stage propulsion system consists of two liquid hydrogen-liquid oxygen engines, performing the critical tasks of final stage acceleration, orbital coasting, and final payload insertion into orbit. This is achieved by establishing mathematical models of the key distributed sensors throughout the rocket and generating sensor data (thrust values, etc.) for the propulsion system based on these models. By determining the optimal noise parameters and navigation state at the current moment, the acquired sensor data is corrected. Through state correction decisions, feedback correction of velocity, position, attitude, and accelerometer zero bias is achieved, the filter noise parameters are optimized, the simulation accuracy of the launch vehicle navigation signal is improved, and inertial measurement data is generated.

[0046] Finally, onboard BIT data is generated based on the self-testing circuit and self-testing software inside the launch vehicle; BIT (Build-In-Time) technology relies on dedicated self-testing circuits and software within launch vehicles and other similar devices. Onboard BIT structures can be broadly categorized into distributed and centralized structures. Distributed BIT structures refer to subsystems or equipment designed with their own fault detection and isolation functions, using inductive reasoning based on test results to determine system functionality. Centralized BIT focuses on aspects not covered by distributed BIT, with the system's central processing unit collecting status information and utilizing system-level testing and diagnostic functions to determine system health. Based on system characteristics and usage patterns, BIT usage modes can be further categorized into power-on BIT, startup BIT, periodic BIT, and continuous BIT. Power-on BIT primarily performs rapid checks before mission execution to determine equipment functionality and readiness for normal use, providing a pass / fail indication. Startup BIT is used for specialized tests, mainly for detecting critical system functionalities and isolating system-level faults. Periodic and continuous BIT continuously or periodically monitor the operational status of various components during mission execution, particularly those affecting safety and mission criticality, primarily for system fault isolation and system performance evaluation.

[0047] Each unit in the control system mainly consists of two parts: functional components and signal processing and control circuits. The signal processing and control circuits of each unit share a similar technical approach, primarily composed of various DSPs, FPGAs, A / D sampling circuits, D / A conversion circuits, bus data communication interface modules, and power supply modules. During BIT design, fault tree modeling is typically performed based on the characteristics of each module, and a targeted BIT detection algorithm is designed based on the fault tree. Based on the characteristics of the control system's unit design, the following method is used to implement BIT testing for such circuits.

[0048] Self-closed-loop test: Utilizing the computing and control resources of digital devices such as DSPs and FPGAs within each functional module, a handshake and status query program is set up inside the circuit for self-testing and mutual testing to complete the self-closed-loop test of the digital signal processing circuit.

[0049] Analog circuit wrap-around test: Electronic switches or relays are set up in the circuit to support the circuit. The signal wrap-around method is used to realize the mutual detection between the D / A output and the A / D sampling circuit, and at the same time, the output characteristics of the power module are tested.

[0050] Online monitoring of operating status: The signal processing circuit's information acquisition channel performs online periodic monitoring of information characterizing the operating status of a single unit, such as operating indication signals, operating voltage, and current. For functional components of the unit, such as the inertial measurement unit (IMU) and servo mechanisms, BIT testing can be performed using test stimuli provided by the unit itself or the natural environment, tailored to the specific functions of the unit. The IMU mainly consists of three accelerometers, three gyroscopes, an I / F circuit board, gyroscope control circuitry, data acquisition and processing circuitry, and frame control and angle measurement circuitry.

[0051] Utilizing local gravitational acceleration in a static base environment and Earth's rotational angular velocity To stimulate the inertial gyroscope and accelerometer, data is obtained over a certain period of time. The internal output values ​​are related to the following formula: , They are inertial navigation systems , , A triaxial accelerometer and a gyroscope The average output value within the range, , These are the acceptable threshold values ​​for the acceleration channel and the angular rate channel, respectively. Furthermore, for a three-autoinertial navigation system, attitude excitation can be obtained by rotating its internal frame. and The output polarity is tested, and the above method can be used to determine whether the inertial measurement unit function is normal and usable.

[0052] The servo mechanism consists of a servo controller and an electromechanical actuator. The servo controller drives the electromechanical actuator to actuate according to angle commands issued by the flight control computer, causing the rocket's aerodynamic control surfaces or engine nozzle to deflect at a given angle. Based on this characteristic, given a specific position in space, the servo controller runs a self-test program to drive the motor to rotate the control surfaces towards that position, determining whether the difference between the measured angular position and the design angle meets the design judgment threshold, thereby determining whether the servo mechanism is operating normally.

[0053] The power-on self-test of the launch vehicle servo system is used to test components such as the servo controller, motor driver, gyroscope, and angle sensor. It adopts a two-loop control scheme, consisting of a gyroscope speed loop and an angle position loop.

[0054] The system works by automatically entering a self-test process after power-on. It moves to the limit position and returns to the zero position within a specified time, and reports that the self-test is normal; otherwise, the self-test times out and anomalies occur.

[0055] When the external platform is stationary, the gyroscope measures the velocity of the mechanism, functioning as a tachometer; the mechanism moves according to the given velocity. When the external platform moves, the gyroscope measures the combined velocity of the mechanism's motion and the external platform's motion, i.e., the inertial space velocity. When the external velocity is opposite to the given velocity, the mechanism's speed increases, and the servo system's self-test time decreases. When the external velocity is in the same direction as the given velocity, the mechanism's speed decreases; the servo system's self-test time increases, and exceeding the specified self-test time causes a power-on self-test anomaly.

[0056] The non-ideal and failure modes of sensors in the structural system are modeled as a combination of "multiplicative-additive-nonlinear operators-random switching".

[0057] The measurement output is written as: in, This is the scaling factor error. Zero bias, It is colored noise. This refers to digital domain noise (quantization jitter / data transmission noise). Indicates saturation / shear nonlinearity. This represents the quantization operator. To describe the entry and exit of various fault types over time, an implicit Markov switching variable is introduced. Controlling the activity of each sub-model, fault parameters Follow Switch or gradually change.

[0058] The scaling factor and zero bias are characterized by a Gaussian-Markov process to depict the slow drift, making and It has adjustable time correlation: in, Zero-mean white noise, Control the drift timescale; when a sudden fault is required, make That's it. Saturation distortion is parameterized by the upper limit of amplitude and the recovery time constant. The sheared signal may exhibit a "flat top + ringing" pattern under impact conditions, which can be addressed by superimposing a decaying sine wave in the saturation region. Description. For the "freezing / stuck" phenomenon, a time-varying gate function is applied to the derivative term, and a hold operator is introduced. For "intermittent packet loss / missing data," a semi-Markov chain is used to generate missing data segments with durations following a log-normal distribution, and sampling clock jitter can be superimposed to simulate spectral leakage and aliasing caused by non-uniform sampling. Non-Gaussian noise environments (such as the long-tailed characteristics of separation shocks) are replaced with Student-t or Laplace distributions, and a mixture model is used. It approximates the "random large deviation".

[0059] ③ Typical Fault Modeling Module The typical fault modeling module is used to simulate typical faults that a rocket may encounter during flight. These typical faults include: inertial navigation system faults, single engine faults, engine thrust reduction, control system faults, structural system faults, and onboard computer faults.

[0060] The typical fault modeling module includes the following sub-modules: The powertrain fault modeling submodule generates key data output by the sensor data generation module to simulate typical engine faults. Rocket engines are the core propulsion system of space launch vehicles. During flight, they may encounter various faults, such as thrust reduction, incomplete combustion, and aerodynamic failures. These faults directly affect the rocket's flight performance and safety. In this embodiment, by modeling each subsystem of the engine, typical engine faults are simulated, thereby providing data support and theoretical basis for fault detection and diagnosis. Typical engine faults specifically include: Single engine failure: When the operating parameters of the combustion chamber are disturbed and the pressure fluctuates significantly, this phenomenon is called unstable combustion. When combustion is unstable, strong engine vibration, mechanical damage to the thrust chamber and other components, and ablation and melting of the injector surface, nozzle, and inside the combustion chamber can all lead to engine failure. When the vibration frequency exceeds 1000Hz, it is called high-frequency unstable combustion, which is the most likely to occur, the most difficult to suppress, and must be addressed. It is characterized by high volatility and high-pressure fluctuations, sometimes reaching 50-100% of the stable pressure, with vibration accelerations ranging from tens to hundreds of grams, and even producing intense flashes and a "sharp whistling" sound in the flame.

[0061] Furthermore, unstable combustion can also cause mechanical damage to the engine. Due to the presence of high-frequency vibrations, the coupling between airflow and mechanical components within the combustion chamber is enhanced, leading to excessive mechanical stress on components such as turbines, pumps, and combustion chamber walls. Components exposed to high-frequency vibrations for extended periods may experience fatigue damage or fracture, increasing maintenance and replacement costs and shortening engine lifespan. Particularly concerning are the combustion chamber walls, nozzles, and other components in contact with the gas; these components are constantly subjected to thermal loads and mechanical shocks from vibrations, potentially resulting in localized wear, cracks, or even melting. In severe cases, this can lead to engine structural failure and ultimately, flight mission failure.

[0062] In summary, unstable combustion, especially high-frequency unstable combustion, not only leads to unstable thrust output of rocket engines, but also causes mechanical fatigue, ablation and damage to various engine components (including turbopumps, nozzles, combustion chambers, sensors, etc.), ultimately leading to the failure of a single engine.

[0063] Furthermore, because turbopumps operate under rapidly changing conditions with short start-up and shutdown cycles, turbines operate at high temperatures, high pressures, and high speeds, and pumps operate in environments with high-pressure, high-speed, flammable, explosive, highly toxic, and strongly corrosive propellants, rotor imbalance can also lead to the failure of a single engine.

[0064] Engine thrust decrease: Engine thrust can be characterized as: in For propellant mass flow rate, For jet velocity, This refers to the nozzle outlet pressure. Due to environmental pressures, For export area. The decrease in thrust is mainly due to and The root cause of abnormal changes can often be traced back to injector blockage or ablation.

[0065] An injector is a special device that introduces and atomizes propellant components into the combustion chamber. Due to its small nozzle diameter and the unique structure formed by processing debris, foreign matter, and flocculent material in the propellant as it flows through the nozzle, it is easily clogged. Nozzle clogging alters the air-fuel mixture in the combustion chamber, leading to extremely high local temperatures. When the clogging is insufficient to change the main system propellant flow rate, the combustion chamber can still operate; that is, the duct remains operational, and the overall air-fuel mixture within the combustion chamber remains unchanged. However, when the clogging causes a change in the main system propellant flow rate, both the overall air-fuel mixture and flow rate in the combustion chamber will change, resulting in a decrease in engine thrust.

[0066] The control system fault modeling submodule is used to simulate faults in rocket servo mechanisms. As the execution unit of rocket flight control, the performance of the rocket servo mechanism directly affects the attitude control accuracy and flight safety. The servo mechanism is often composed of a motor, transmission, valve body, and actuator, making the system complex and highly coupled.

[0067] In actual service, faults such as sensor drift, valve core jamming, and torque reduction occur frequently. If these faults are not identified and accurately modeled in a timely manner, they can lead to control failure or even serious accidents. Therefore, a fault mathematical model of the rocket servo mechanism is constructed to simulate the faults of the rocket servo mechanism.

[0068] The inertial navigation system fault modeling submodule is used to simulate faults in the inertial navigation system of a launch vehicle. The errors of inertial navigation systems are mainly caused by sensor errors (such as zero bias, scale factor error, and noise of accelerometers and gyroscopes), initial alignment errors (such as inaccuracies in initial position, velocity, and attitude), calculation errors (such as numerical integration and algorithm errors), and environmental interference (such as external factors such as vibration and temperature changes).

[0069] The specific process of modeling inertial navigation system faults based on the working principle and error propagation mechanism of inertial navigation systems is as follows: First, define the state vector. for: in, , , These are represented as position error, velocity error, and attitude error, respectively. Next, based on the working principle of the inertial navigation system, the error dynamics equation is derived: in, The system matrix describes the propagation characteristics of the error. The noise input matrix, This represents process noise. Then, based on the physical characteristics of the navigation system, the system matrix is ​​calculated. The individual submatrices: in, , , Describe the propagation of position error. , , Describe the propagation of velocity error. , , The propagation of attitude error is described. Furthermore, error models for the accelerometer and gyroscope are established, including zero bias, scale factor error, and noise. The accelerometer error can be expressed as: in, For the measurement error of the launch vehicle accelerometer, Zero bias, This is the scale factor error matrix. This refers to accelerometer noise. Gyroscope error can be expressed as: in, For the measurement error of the launch vehicle gyroscope, Zero bias, This is the scale factor error matrix. This is to account for gyroscope noise. Finally, the error model is simulated using the numerical simulation tool MATLAB / Simulink, and the trend of error variation over time is analyzed to evaluate the performance of the navigation system.

[0070] The navigation and control system fault modeling submodule is used to simulate the control surface failure of a launch vehicle. Failure of launch vehicle control surfaces (such as rudder surfaces) is a significant type of navigation and control system malfunction, potentially leading to attitude instability or deviation from the flight trajectory. The specific process for modeling this type of failure is as follows: First, control surface failure typically includes several modes, such as complete failure (control surface stuck or unable to move), partial failure (control surface action range limited or response speed reduced), drift failure (control surface deflection without command), and delay failure (time delay in control surface response to command). Next, a control surface dynamic model is defined. Based on the physical characteristics of the control surface, the dynamic equations are established: in, To control the surface deflection angle, For control input (instructions). To control surface gain, The time constant reflects the response speed of the control surface. Next, based on the fault type, a complete failure, partial failure, drift failure, or delayed failure model is selected. In the case of complete failure, the control surface deflection angle is fixed at a certain value. : When partial failure occurs, the range of control surface deflection angles is limited: in, and These represent the minimum and maximum deflection angles of the control surface, respectively. In the event of drift failure, the control surface deflects without command: in, The initial deflection angle, This refers to the drift rate. In the event of a delay failure, the control surface response command experiences a time delay. in, This is the delay time. Then, based on Euler's equations, a rocket attitude dynamics model is established: in, Here is the rotational inertia matrix of the rocket. It is the angular velocity vector. The torque generated for the control surface satisfies , Let be the position vector from the control surface to the rocket's center of mass. The aerodynamic forces generated by the control surfaces are then analyzed. Finally, a control surface failure model is constructed and simulated using the numerical simulation tools MATLAB / Simulink, facilitating subsequent analysis of the impact of control surface failure on rocket attitude and trajectory.

[0071] The onboard computer fault modeling submodule is used to generate hardware circuit and software operating status data of the launch vehicle's engine and to perform fault simulation based on the generated status data. As a crucial control unit of a rocket, the rocket engine's proper functioning is essential for the success of its flight mission. This module provides comprehensive monitoring and diagnostics of the rocket engine from multiple dimensions, including hardware circuitry and software operation status. On the hardware side, a specially designed hardware monitoring circuit monitors key parameters such as power supply voltage, temperature, and clock signal in real time. When abnormal fluctuations in power supply voltage, excessively high temperatures, or clock signal loss occur, an alarm can be issued promptly, indicating a potential hardware failure.

[0072] (2) Online intelligent diagnostic system The online intelligent diagnostic system is used to monitor the onboard status and perform online fault diagnosis based on key data output by the spacecraft simulation and fault data generation system. The key data includes high-fidelity sensor data generation and highly reliable fault simulation data, with the fault simulation data mainly consisting of sensor data. like Figure 6 As shown, the online intelligent diagnostic system includes the following sub-modules: ① Self-health assessment module The autonomous health assessment module is used to preprocess the sensor data output by the spacecraft simulation and fault data generation system, and to monitor the status based on the preprocessed standardized data, and to issue alarms for abnormal states; the sensor data includes inertial measurement data, engine thrust sensor data, BIT information and servo self-test data, etc. Specifically, firstly, the sensor data from each channel of the space launch vehicle needs to be standardized to obtain standardized data. In this module, linear function normalization, wavelet denoising, and Lagrange interpolation methods are used to perform standardized preprocessing on the raw sensor data from each channel of the space launch vehicle. The standardized preprocessing specifically includes the following steps: Step 1: For raw sensor data with uneven distribution, use the Lagrange polynomial interpolation algorithm to fill in the missing parameters; Step 2: Denoise the collected incomplete rocket sensor sample data based on wavelet threshold denoising; Wavelet thresholding denoising is a classic signal denoising method. It divides the signal into frequency bands of varying widths based on wavelet transform, and then denoises the wavelet coefficients in different bands through thresholding. In wavelet thresholding denoising, the setting of the threshold function is crucial to the denoising effect. Traditional threshold functions include hard thresholding and soft thresholding. The expression for the hard thresholding function is: in, After processing with the threshold function, the first The first in scale Wavelet coefficients, Before threshold function processing, the first The first in scale Wavelet coefficients, For threshold; Hard thresholding functions are nonlinear and discontinuous functions, offering good signal binarization but lacking in smoothness. Soft thresholding functions are nonlinear, continuous, and differentiable functions, providing good signal smoothing but losing some signal characteristics at the threshold. To overcome the shortcomings of traditional thresholding functions, they are constantly being improved. The most common compromise between hard and soft thresholding functions enhances denoising by introducing an adjustment factor, as shown in the following expression: in, As an adjustment factor, its value ranges from between.

[0073] However, this threshold function not only fails to solve the distortion problem when the absolute value of the wavelet coefficients is less than the threshold, but also fails to describe the decay process of the wavelet coefficients well.

[0074] Therefore, this application proposes a new wavelet threshold function by introducing exponential functions with different bases, the function expression of which is as follows: in, As a regulating factor, and always ensuring Regulatory factors The principle for determining the value of the adjustment factor is as follows: when the signal-to-noise ratio of the noisy signal is small, the adjustment factor... A larger value helps to remove more noise wavelet coefficients; when the signal-to-noise ratio of the noisy signal is high, the adjustment factor... Taking smaller values ​​helps preserve wavelet coefficients for weaker signals. Generally, a value of [value missing] is used. This approach can achieve a good denoising effect. In this embodiment, when the optimal adjustment factor needs to be determined for different signals, the signal-to-noise ratio function is used as the fitness function, and the particle swarm optimization algorithm is employed to find the optimal value, with the adjustment factor obtained when the signal-to-noise ratio is maximized being the optimal value.

[0075] The signal-to-noise ratio function is defined as follows: in, For signal-to-noise ratio, This represents the signal duration.

[0076] The threshold selected is a fixed threshold that is fast to calculate and widely applicable. The definition of the fixed threshold is as follows: in, The standard deviation of the wavelet coefficients after thresholding; like Figures 7-8 As shown, the new threshold function proposed in this application solves the problems of discontinuity at the threshold in the hard threshold function and fixed deviation in the soft threshold function.

[0077] Next, in this module, we utilize an LSTM neural network based on a fusion attention mechanism to predict sensor data. LSTM neural networks can effectively solve the gradient vanishing and gradient exploding problems of traditional RNNs during training on long sequences. However, in some tasks, the fixed-time-step attention mechanism of LSTM may be insufficient to handle long-range dependencies. Introducing an attention mechanism can help LSTM more flexibly focus on key information at different time steps when processing input sequences, rather than simply relying on fixed sequential information.

[0078] The process of using an LSTM neural network to predict sensor state data is as follows: rocket sensor data is first sent to the forget gate to clear previously useless cell states; then the cell states are updated through the input gate; finally, the cell states (sensor state prediction data) are output through the output gate.

[0079] While LSTMs are effective at capturing long-term dependencies in time series, they still have limitations, particularly in handling these dependencies. The model may be insufficient in selecting data at key time steps. For example, when faced with complex time series data, LSTM models may rely on fixed-time-step input information, failing to flexibly focus on certain key moments in the sequence. This can lead to the neglect of important information in fault detection tasks, reducing the model's predictive power and accuracy. Therefore, this application utilizes an LSTM neural network to first encode the input sequence, generating a series of hidden states. Then, an attention mechanism is used to calculate the importance of each hidden state, typically based on the similarity between the current hidden state and the encoder's hidden states. The calculated attention weights are then used to weight and sum the LSTM hidden states to produce the final context vector.

[0080] The attention mechanism assigns different weights to the input at each time step, enabling LSTM to automatically focus on the parts of the sequence more relevant to the current task, rather than relying solely on fixed time steps. This makes the model more flexible in handling long sequences and better able to capture long-term dependencies. The attention mechanism simulates human behavior when processing information, selectively focusing on certain parts while ignoring irrelevant information. Specifically, the model calculates attention weights for the hidden states at other time steps based on the current hidden state, and then sums these weights across the information at different time steps. For each hidden state at each time step of the LSTM, a query matrix, a key matrix, and a value matrix are defined. These matrices are extracted from the LSTM output and used to calculate the relative importance of the input at each time step. The attention weights for each time step are obtained by calculating the similarity between the query matrix and the key matrix; these weights reflect the dependence of the current time step on other time steps. Then, based on the calculated attention weights, the hidden states at different time steps are summed across the matrix to generate the final context vector, which contains the most important information at the current time step. The mathematical model of the attention mechanism is as follows: in, Query the query matrix separately Key matrix AND-value matrix The weight matrix, Represents the attention score matrix; Multi-head attention applies the attention mechanism to multiple attention heads, each capturing different parts of the input sequence. The results from multiple attention heads are concatenated or weighted averaged to obtain the final attention output. As an improvement on the traditional attention mechanism, multi-head attention maps each set of inputs to different sub-representation spaces, allowing the model to focus on different positions in different sub-representation spaces. This avoids drastic performance degradation caused by bias in the weight matrix while more comprehensively exploring multi-dimensional correlations between time series. The entire computation process can be represented as: in, It is the first Attention score matrix under single-head attention mechanism The multi-head attention score matrix, To output the transformation matrix, This is a matrix concatenation function.

[0081] Furthermore, rocket sensor data health assessment is not a simple classification problem. False alarms and missed alarms can occur during the health assessment process. To reflect the probability of these two situations occurring, it is necessary to calculate the accuracy of the model's detection. With recall rate : in, This represents the number of normal samples that were correctly detected as normal samples. This represents the number of abnormal samples that were mistakenly identified as normal samples. This indicates the number of normal samples that were falsely detected as abnormal samples.

[0082] This application uses an LSTM neural network with a fused attention mechanism to predict sensor state data, and uses the F1 evaluation metric to evaluate the performance of the constructed LSTM neural network with a fused attention mechanism. The formula for calculating the F1 evaluation index is as follows: (12) To further improve prediction accuracy, rocket model information is incorporated when training the neural network with an attention mechanism. The physical model is embedded into the loss function of the neural network in the form of soft constraints, thereby constructing a physical information neural network.

[0083] For the power system, its physical losses Composed of multiple sub-items, the network predicts the oxidizer and fuel flow rates. Should be related to combustion chamber pressure and throat area Satisfy the nozzle flow rate formula: in It can be calculated based on the predicted combustion chamber temperature and mixture ratio; The thrust predicted by the network should be consistent with the predicted flow rate and nozzle parameters: The physical constraints of the inertial navigation system (INS) originate from the rigid body dynamics equations of a rocket. The network input is a time-series state, and its output is acceleration and angular velocity. The differentials of the outputs satisfy the Newton-Euler equations: Among them, thrust and torque Aerodynamic force can be provided by the output of the neural network of the power system. and torque Calculated using a simplified aerodynamic model, the total loss predicted by the inertial navigation system is then determined. for: The physical constraints predicted by the control system can be derived from the mechanistic model of the actuator, using simplified flow equations and force balance equations for the servo hydraulic system as constraints: in, It is the predicted servo valve flow rate. It is the predicted valve core displacement. It is the predicted oil pressure. It is the predicted rudder deflection angle.

[0084] The physical constraints predicted for structural systems mainly come from the structural dynamics equations: in These are the mass, damping, and stiffness matrices, respectively.

[0085] Construct the physical loss term, requiring the predicted response to satisfy the kinetic equation: Among them, incentive Predictions can be derived from dynamic systems and aerodynamic loads; differential terms Obtained through automatic differentiation.

[0086] Furthermore, in deep learning model training, hyperparameter tuning is a crucial step in improving model performance. Especially with the massive scale of rocket sensor data, neural network models typically require complex structures and vast parameter spaces. As the model size increases, the number and space of hyperparameters also expand significantly. Even minor hyperparameter adjustments can have a huge impact on the model's training effectiveness, convergence speed, and final performance. Therefore, hyperparameter tuning is essential for improving training efficiency and accelerating convergence when training neural network models. In practical applications, rocket sensor data is often massive, high-dimensional, and complex, posing a significant challenge to neural network training. Large-scale model training requires substantial computational resources and time, and over-reliance on traditional methods such as manual tuning and grid search can become extremely inefficient when the hyperparameter space is large. Therefore, hyperparameter optimization becomes a critical step.

[0087] To improve tuning efficiency, save computational resources, and accelerate model training, this application employs a Bayesian optimization method for hyperparameter tuning of an LSTM neural network model with an attention fusion mechanism.

[0088] Bayesian optimization is an optimization method for expensive or time-consuming functions, particularly suitable for optimization in large hyperparameter spaces. Bayesian optimization builds a probabilistic model by tracking past evaluation results, then uses the surrogate model to infer which hyperparameter combinations are likely optimal, and selects these combinations for testing in future evaluations. The process of Bayesian optimization is as follows: Figure 9 As shown.

[0089] Combination Figure 9 In Bayesian optimization, the surrogate model is typically a Gaussian process, providing an efficient way to estimate the behavior of a function, especially when there is insufficient actual evaluation data. The main advantage of Bayesian optimization lies in its ability to balance exploration and exploitation; it can search in underexplored hyperparameter regions while further optimizing known high-performing regions, achieving more efficient hyperparameter tuning. Bayesian optimization methods form a probabilistic model by tracking past evaluation results, mapping hyperparameters to the score probabilities of the objective function. This model is called the "surrogate" of the objective function and is represented as... Since the surrogate is easier to optimize than the objective function, Bayesian optimization methods optimize hyperparameters based on their performance on the surrogate model.

[0090] First, the cross-entropy loss function is defined as the objective function for Bayesian optimization. Then, Gaussian process regression is used to model the objective function, thereby predicting the performance of different hyperparameter combinations without actual evaluation. Gaussian processes assume that function values ​​follow a normal distribution, and a multidimensional Gaussian distribution is established to describe the latent values ​​of the objective function. Given a combination of hyperparameters, the surrogate model predicts new hyperparameter combinations based on past observations. The core of Gaussian processes lies in describing the similarity between different inputs through a kernel function. The goal of Gaussian process regression is to infer the distribution of objective function values ​​using historical data, thereby predicting the performance of other hyperparameter combinations without actual evaluation. In Gaussian processes, any linear combination of random variables follows a normal distribution, assuming that similar inputs produce similar outputs. A function is constructed based on this assumption and is completely determined by its expected value and covariance function, thus inheriting many properties of the normal distribution. The formula for the Gaussian distribution is: in, It's about variables. The probability density function of the normal distribution. Let them represent two different hyperparameter vectors over the continuous domain of the Gaussian process. Represents the mean function, The kernel function that determines the similarity between function values ​​is represented as: The training process for Gaussian process parameters is as follows: First, several sets of hyperparameters of the neural network model are uniformly selected. Input the probability density function to obtain the corresponding output. The output results follow a mean of 0 and a covariance matrix of... The multivariate Gaussian distribution is used as the prior dataset for the Gaussian process. Based on the established function model, new sampling points are calculated. and the corresponding function output value: in, .

[0091] After fitting the prior data using a Gaussian process, we obtain information about... A probability distribution of the values ​​is used to estimate the values ​​of new sampling points. Then, the newly collected samples are used to update the posterior through the objective function. This process is repeated until an acceptable optimization point of the objective function is reached, and the optimized model hyperparameters are output.

[0092] Select the optimal point directly from the proxy model Evaluation is typically costly. In each iteration of Bayesian optimization, a collection function α is usually used to obtain the next sample point for evaluation. The collection function is composed of the posterior distribution of the observed data; maximizing this function selects the next sample point to be evaluated. The PI acquisition function maximizes the probability that a new sample improves upon the current objective function's maximum value. Let... If the current optimal value is obtained, then better hyperparameters can be obtained by maximizing the following acquisition function: Acquisition Function for: in, Represents the normal cumulative distribution function. It is a random parameter used to balance exploration and exploitation. When the value is large, PIs tend to explore unknown information. When the information is small, PIs tend to utilize known information; Bayesian optimization, by effectively balancing exploration and utilization, can efficiently search and optimize the hyperparameters of LSTM neural networks with fused attention mechanisms within a vast and high-dimensional hyperparameter space. By introducing techniques such as Gaussian process regression, kernel function selection, and acquisition function maximization, Bayesian optimization significantly improves the efficiency and accuracy of hyperparameter tuning, reduces training time, and enhances model performance. This process not only accelerates the training of deep learning models for rocket sensor data processing but also provides an efficient optimization path for fault detection and health assessment tasks.

[0093] Finally, by calculating the residual between the LSTM predicted signal and the actual observed signal, and combining this with the Z-score normalization method to set a dynamic threshold, high-confidence identification of anomalous signal channels can be achieved. Specifically, the predicted output value of the LSTM neural network model with an attention mechanism is subtracted from the actual value of the rocket sensor data to obtain a residual sequence. By analyzing the residual sequence, anomaly thresholds are determined, thereby enabling timely detection of anomalous data.

[0094] The basic window thresholding method calculates the threshold by linearly adding the sequence mean and variance. However, rocket sensor signals exhibit diverse anomaly patterns, and this method, being too simplistic, cannot effectively identify all anomalies. Furthermore, due to the long-term, phased oscillations in rocket anomaly signals, the conventional window thresholding method cannot retain long-term data trends, failing to accurately detect anomalies in their later stages. Additionally, continuous amplitude jumps in rocket sensor data significantly impact the average residual data within the window, leading to frequent false alarms. Moreover, the window size and variance weights in the basic window thresholding method lack mechanistic model support, making it difficult to balance accuracy and recall, thus failing to meet the anomaly detection requirements of complex signals. Therefore, this application employs a dynamic thresholding method with secondary scaling. This method calculates a dynamically changing threshold based on the residual sequence; the threshold at adjacent sampling times is not the same but is continuously updated according to the residual changes. Furthermore, this algorithm adopts a secondary judgment strategy for suspected anomaly data, reducing false alarms and improving detection accuracy to adapt to various detection scenarios. The dynamic thresholding method specifically includes: First, the data window length used for threshold updates is defined as follows: Calculate the residual vector The mean and variance of all elements in the set are determined before... Abnormal threshold at each sampling time : (twenty four) in, This is the variance weighting coefficient.

[0095] Secondly, the residual data that is detected as normal is denoted as a vector. If vector Insufficient number of elements Then, following the previous step, perform the following steps... Perform the test and add the normal residual data to End. Repeat the above steps until... The number of elements in the middle is greater than or equal to .

[0096] Finally, update the sampling time. threshold : (25) in, The elements in are vectors end Each element.

[0097] If the residual value Not exceeding the threshold Then Add to vector Finally, if the threshold is exceeded, a second detection is performed to prevent false alarms.

[0098] Construct normal data sequences from rocket sensor data detected as normal. Calculate vector end The signal peak value at each sampling time is used to perform secondary detection by changing the threshold. (26) in, The threshold value for secondary detection. These are the actual values ​​of the rocket sensor data to be tested a second time.

[0099] Compared to the basic window threshold method, the dynamic scaling threshold method can effectively improve detection accuracy while ensuring recall, thereby increasing... value.

[0100] ② Autonomous Fault Identification Module The autonomous fault identification module is used to analyze and identify the received abnormal status alarm information and obtain the fault classification results; To ensure the successful execution of the launch mission, the system needs to conduct fault diagnosis on key components of the control system, including servo mechanisms, inertial measurement units, and the rocket's engine (system hardware). This allows for the timely development of fault handling measures, and the diagnostic information is fed back to the onboard computer to ensure efficient and coordinated diagnosis at the component and system levels. Once the rocket enters the launch phase, online fault diagnosis will be performed. The onboard computer will utilize this system to collect information from various components and systems, employing a combination of data and modeling to diagnose control system faults and handle them at the system level until the payload successfully enters orbit, thus ensuring the safety and stability of every stage of the mission.

[0101] Finally, the fault diagnosis results of the inertial navigation system and the power system obtained based on the fault identification in this module are as follows: Figures 11-12 As shown.

[0102] ③ Fault handling decision module The fault handling decision module determines the fault index vector based on the fault classification results, and maps the fault index vector into fault handling instructions through a decision mapping function.

[0103] In the overall rocket health management system, the control system, structural system, and propulsion system are responsible for the core tasks of attitude control, load support, and propulsion, respectively. The rocket's control system mainly includes servo mechanisms, inertial measurement units, and system hardware; these components directly determine the rocket's attitude control and flight stability. The structural system encompasses the rocket body, connecting components, and load support structures; its health status relates to the rocket's overall strength and vibration resistance. The propulsion system consists of the engine, fuel tanks, and fuel supply system, directly providing the rocket with the required propulsion force. Therefore, high-precision sensors are deployed at key parts of the rocket to collect real-time data on current, voltage, torque, acceleration, strain, temperature, pressure, fuel flow, and thrust, forming the basis for rocket health monitoring. Due to environmental disturbances, mechanical wear, temperature changes, and uncertainties during combustion, various subsystems are susceptible to different types of failures during high-speed flight. Assuming each system is equipped with high-precision sensors and undergoes real-time monitoring through preprocessing, data fusion, and fault diagnosis algorithms, the system can obtain information such as the current fault type and severity within milliseconds. At this point, mapping different fault information to corresponding fault handling instructions becomes a key aspect of ensuring the overall safety and efficient response of the rocket.

[0104] Specifically, mapping different fault information to corresponding fault handling instructions includes the following steps: First, multi-source data is collected based on high-precision sensors on various key components; the multi-source data includes: servo mechanism sensor data, including motor current. ,Voltage Torque ,angle and temperature Inertial measurement unit sensor data, including acceleration angular velocity and attitude angle System hardware sensor data, including monitoring power supply voltage. and communication delay .

[0105] Since the original signal usually contains noise, an exponentially weighted moving average (EWMA) is used for low-pass filtering to improve data quality. Its mathematical expression is: in The original sampled signal, This is the filtered signal. This is a smoothing factor (typically ranging from 0.1 to 0.3). Furthermore, to eliminate the influence of different sensors' dimensions, the system normalizes the data from each channel.

[0106] After filtering and normalization, the data undergoes multi-sensor data fusion processing, and the state is estimated in real time using Kalman Filter (KF).

[0107] Based on the state estimation, the system compares the difference between the actual measurements and the theoretical model output as the response error. Similarly, for the inertial measurement unit, the system calculates the cumulative deviation of acceleration and angular velocity. As an important parameter for judging inertial drift, and to generate fault indicators. For system hardware, fault indicators are generated by monitoring power supply fluctuations and communication delays. In summary, the fault index vector of the control system can be written as: Next, a decision mapping function is used to map fault indicators and real-time status information to specific processing instructions. Let the current control system state be... (Including the fused states of each sensor), the system design parameters and safety constraints are denoted as... The decision mapping function is: in, This is used to process commands generated by the control system. For example, in a servo mechanism, when a fault indicator... When a certain level is reached, the system will calculate the relative error: And process them according to this classification: when When the command is given to "adjust PID parameters" or "increase current compensation", the command will be issued. When the system switches to redundant servo channels, and when In such cases, the system issues commands to "activate the backup servo system" or "emergency shutdown protection." These commands are transmitted to the servo drive module and flight controller via internal communication links, ensuring that the rocket maintains high attitude control accuracy even in the event of a malfunction.

[0108] During rocket launch and flight, the structural system constantly bears the heavy responsibility of withstanding enormous loads and coping with high-speed dynamic environments. Therefore, it is highly susceptible to stress concentration, vibration resonance, localized fatigue, and even fracture under extreme conditions. To ensure the overall safety of the rocket and mission success, real-time monitoring, fault diagnosis, and automated handling of the rocket's structural system are essential. This module, based on high-precision multi-sensor data acquisition, achieves real-time assessment of the rocket's structural health status through multi-level preprocessing and fusion of sensor signals. Utilizing advanced physical models and data-driven algorithms, it constructs fault indicators and decision mapping functions, rapidly transforming abnormal states into emergency response commands. This ensures that the rocket can quickly take effective remedial measures when a fault occurs, thereby achieving intelligent and automated fault management of the system. After the fault detection phase of the rocket's structural system is completed, the system has obtained fault indicator data for each key component. These data are typically represented in vector form, denoted as: Each of them A fault indicator representing a certain area or component, which may be derived from the fusion of multiple monitoring data such as strain, vibration, and temperature.

[0109] Based on these fault indicators, corresponding handling instructions are issued according to different fault types and severity levels. First, a comprehensive fault indicator is introduced. To quantify the overall failure risk, it is defined as the weighted sum of all individual failure indicators: in, Each indicator is assigned a weight based on its importance; these weights can be determined based on historical data, experimental analysis, or expert experience. This comprehensive index reflects the current health status of the rocket's structural system, providing a basis for subsequent decision mapping.

[0110] To map fault detection results to specific handling instructions, this module designs a set of decision mapping functions. This function accepts fault detection vectors. or comprehensive indicators As input, the output processing instruction vector : in, This represents the system design parameters, adjustment coefficients, and safety constraints. The entire decision mapping process aims to quantify fault information and transform it into actual handling actions through a certain mathematical model, such as initiating local hardening, adjusting load distribution, activating backup protection systems, or directly implementing emergency load reduction.

[0111] To achieve the aforementioned decision mapping, this module proposes an instruction generation model based on piecewise functions and continuous mapping. It is assumed that for a single fault index... Corresponding handling instructions It can be determined using the following formula: In this formula, and The first The low-risk and high-risk thresholds for each fault indicator, when Less than When the system considers the indicator to be in a normal state, no action is issued; when In Within the specified interval, the system employs a relatively mild processing measure, with its instructions processed through a linear function. Generate; and when Exceed At this time, the system considers there to be a serious risk of failure and requires emergency response measures, and its instruction is: The above parameters can all be calibrated based on actual conditions using experimental or simulation data, thereby enabling dynamic adjustment of the command output.

[0112] In real-world systems, fault indicators are often not singular but multi-dimensional vectors. Therefore, a global instruction vector can be defined. Its generation depends not only on a single indicator, but also on the overall distribution of fault detection results. Therefore, a matrix-form decision mapping model can be constructed: in , is the decision gain matrix, where each row represents the response coefficient of a certain processing module to various fault indicators. This represents the baseline value vector for each indicator. When certain fault indicators exceed the set baseline values, the decision gain matrix amplifies the corresponding deviation and generates corresponding control commands. For example, in cases of loose connection components or localized crack propagation, setting higher weights can make even minor deviations significantly reflected in the command output, thereby prompting the system to take timely measures to reinforce or switch redundant structures.

[0113] In more complex scenarios, to address the nonlinear relationships between multiple variables in a rocket structural system, decision mapping functions often take the form of nonlinear functions. Based on this consideration, this module introduces a mapping model based on the Sigmoid function, which can achieve smooth nonlinear transformations. Specifically, the generation function for a single instruction is defined as follows: in, This represents the maximum handling instruction amplitude corresponding to this fault indicator. For the Sigmoid function, It is the slope factor, and This is the intermediate threshold. When... much smaller hour, Approaching 0, instruction Almost zero; when Exceed Then, instructions Rising rapidly, approaching The value of . This nonlinear mapping can achieve a smooth transition of instruction output, avoid the risk of sudden changes that may be caused by piecewise functions, and has good robustness and adaptability.

[0114] Furthermore, to comprehensively consider the interactive effects of multiple fault indicators, a fuzzy logic system is introduced to fuzzify the fault detection results. Assume that for each fault indicator... Define a membership function This indicates the degree to which a function belongs to a fuzzy set such as "normal", "minor fault", "moderate fault", or "serious fault", and can be expressed using a Sigmoid membership function as follows: in, Indicates different risk levels (e.g.) It is mild. Moderate. (For severe cases), is the slope parameter. This is the corresponding intermediate threshold.

[0115] Next, based on the membership degrees of each fuzzy set, a weighted average method is used to determine the final instruction output: in, This represents the preset handling instruction amplitude for the corresponding risk level. Through this fuzzy logic processing method, the system can achieve more flexible and precise instruction decisions when facing complex and uncertain fault states.

[0116] In practical applications, rocket structural systems often have multi-level protection mechanisms. For example, in a low-risk state, it is only necessary to log, alert the monitoring center, and make minor adjustments to flight parameters; in a medium-risk state, the system needs to proactively take measures such as local reinforcement, adjusting load distribution, or activating local backup structures; and in a high-risk state, emergency response procedures must be initiated immediately, such as forcibly reducing the load, disconnecting the linkage of the faulty part, or activating the emergency downgrade mode of the entire system.

[0117] Furthermore, to ensure the rapid and accurate transmission of disposal commands throughout the rocket control system, redundant communication links and a real-time operating system (RTOS) are employed. After generation, disposal commands are transmitted to execution modules, such as structural stiffening actuators, load distribution adjustment modules, and emergency load shedding devices, via a high-speed data bus and dedicated communication protocols. To guarantee communication reliability, error checking and multipath transmission technologies are used during transmission to ensure that commands are not delayed or lost due to communication failures. Simultaneously, to meet safety requirements, real-time monitoring and self-diagnostic mechanisms are embedded throughout the entire command generation and transmission process. These mechanisms continuously monitor the system status and trigger backup plans upon detecting anomalies, ensuring that the entire rocket structural system can achieve multi-level protection and immediate response in the event of a fault.

[0118] In summary, based on fault detection results, this module utilizes a multi-level, piecewise, and nonlinear mapping model to convert multi-dimensional fault indicators into specific handling instructions. This is achieved by constructing a comprehensive fault indicator system. This scheme, employing its piecewise mapping function, nonlinear decision mapping based on the Sigmoid function and fuzzy logic, and real-time parameter correction using an adaptive control algorithm, achieves a closed-loop control process from fault detection to handling command generation. Mathematically, the decision mapping function can be summarized into the following form: Among them, each parameter matrix With threshold All results were obtained through offline data analysis and online adaptive correction. This formula not only reflects the tiered response strategy under different risk levels, but also takes into account the system's need for a smooth transition between low-risk, medium-risk, and high-risk states.

[0119] The propulsion system is the core of rocket propulsion, and its main components include the liquid-fueled rocket engine, fuel tank, and fuel supply system. Propulsion system failures may manifest as insufficient engine thrust, unstable combustion, abnormal combustion chamber temperature, or abnormal fuel supply. Intelligent fault handling systems for rocket propulsion systems need to analyze fault types in real time under complex flight environments and generate corresponding processing commands. Because the propulsion system involves multiple strongly coupled subsystems, such as the combustion chamber, turbopump, propellant supply system, and nozzle, fault handling must possess multi-level decision-making capabilities to ensure the optimal recovery strategy for the system under different fault conditions. Based on Model Predictive Control (MPC), Deep Reinforcement Learning (DRL), and adaptive control theory, high-precision sensors are deployed on key components such as the engine, fuel supply system, combustion chamber, and ignition system to collect real-time data on temperature, pressure, fuel flow rate, thrust, combustion chamber pressure, and combustion temperature, forming the basis for monitoring the status of liquid-fueled rocket propulsion systems. Based on the fault types reported by the front-end system, the deviations and fault indicators of each key parameter are calculated according to the real-time collected data and the state estimation model. Corresponding processing instructions are then generated according to the designed decision mapping function. For example, when insufficient engine thrust is detected, a "restart engine" instruction is output based on a comprehensive judgment. The entire technical implementation process mainly includes data acquisition and preprocessing, state estimation and model identification, error quantification and fault indicator generation, decision mapping and instruction generation, and hierarchical response and issuance of instructions.

[0120] During the data acquisition phase, key sensors of critical components in the powertrain system acquire raw signals at high frequency. These signals include combustion temperature, combustion chamber pressure, nozzle outlet thrust, and fuel supply flow rate within the engine combustion chamber. Due to environmental noise, sensor drift, and digital quantization errors during actual sampling, preprocessing of the raw data is necessary. Therefore, the system employs an exponentially weighted moving average method to perform low-pass filtering on the raw sampled signals. Subsequently, to eliminate the influence of different dimensions among the sensors, the filtered data is normalized. A Kalman filter-based state estimation method is then used to estimate the states of the powertrain system sensors. The estimation process is similar to that of the servo mechanism and will not be elaborated further here.

[0121] (3) Online intelligent planning system The online intelligent planning module is used to re-divide the launch trajectory mission area and re-plan the launch trajectory and control law when the launch vehicle malfunctions, and adjust the launch vehicle's flight attitude and orbit. like Figure 12 As shown, the online intelligent planning system includes the following sub-modules: ① Launch orbit mission reconfiguration module The launch trajectory reconfiguration module is used to determine the reachability domain of the launch vehicle in an abnormal state when a launch vehicle malfunctions. The rocket's thrust system may be damaged or malfunction during flight, causing the rocket to deviate from its intended trajectory. To address the abnormal launch trajectory states during launch vehicle flight, this application introduces random variables to characterize the uncertainties caused by sudden rocket failures and predicts the rocket's future orbital reachability domain based on the state transition tensor method. Furthermore, based on the reachability domain information under abnormal rocket conditions, combined with the apogee altitude deviation, orbital inclination deviation, and different mission conditions, the final reachability range under abnormal conditions is determined, thereby assessing the rocket's orbital reachability and enabling launch trajectory mission reconfiguration. This allows the rocket to achieve optimal capabilities while ensuring safety.

[0122] Furthermore, in order to calculate the reachability domain of the rocket under abnormal conditions, we first explain the actual impact of the failure on the propulsion system after the failure occurs, which is specifically divided into three categories: 1) Thrust loss A slight decrease in overall thrust, an unexpectedly large decrease in overall thrust, complete loss of thrust, and thruster explosions can all directly result in a loss of current and future thrust. If the thrust loss has a minor impact on payload capacity, iterative guidance adjustments to the injection point and flight trajectory can still deliver the payload to the target orbit. However, if the impact is significant, flying towards the target orbit will cause premature propellant depletion. Whether an orbit can be formed in this case depends on the state at the moment of depletion, and it may also lead to rocket crashes.

[0123] For a certain type of fault that causes thrust loss, the thrust loss will be... Modeled as random variables as follows: in, Let be a three-dimensional random variable that follows a normal distribution; It is the expected value of a normal distribution; The standard deviation is the normal distribution. 2) Fuel loss Minor fuel leaks, fuel leaks without explosions, thruster explosions, and stage separation failures can all directly result in fuel loss. Fuel loss primarily affects the rocket's payload capacity; when fuel loss is excessive, the remaining fuel may be insufficient to support the rocket in completing its intended mission.

[0124] For a certain type of fault that causes fuel loss, the fuel loss will be... Model as random variables: in, For random variables that satisfy a normal distribution; It is the expected value of a normal distribution; The standard deviation is the normal distribution. 3) Launch trajectory error While a stage thruster failing to ignite at the scheduled time but successfully igniting again, or a booster failing to separate at the scheduled time but successfully separating again, may not directly result in thrust or fuel loss, it will cause the rocket's current position to deviate from the planned launch trajectory, causing launch trajectory error. Furthermore, other types of malfunctions can also lead to significant unexpected errors in the launch trajectory. Even if thrust and fuel are instantly restored, the launch vehicle still requires additional fuel to compensate for launch trajectory errors and ensure tracking of the planned launch trajectory.

[0125] The impact of the malfunction on the launch vehicle's launch trajectory can be directly quantified and calculated using the current position of the rocket and the current position of the planned launch trajectory: in, This is the rocket's current position vector; The current position vector for planning the launch trajectory; This refers to the launch trajectory error.

[0126] By introducing random variables to quantify the uncertainties caused by different faults, the differential equation of the dynamic system of the launch vehicle containing uncertainties can be obtained as follows: in, for The actual state quantity of a 3D dynamic system; This is the control input for the dynamic system, and the input parameter is calculated by the system based on the navigation algorithm (such as the Kalman filter algorithm) and the guidance algorithm; This is the noise gain matrix; This indicates the acceleration uncertainty caused by the fault.

[0127] In addition, the system's actual navigation status The actual state of the system Confirmed result: The true state of the system under the influence of uncertainty and navigation state quantities All of these will deviate from the expected value, causing the nonlinear system to no longer follow a specific trajectory, but instead move within a trajectory envelope. To measure the impact of uncertainty on a nonlinear system, the true state is defined. The state deviation is Navigation state variables The state deviation is ,but and The envelope ranges of the real state trajectory and the navigation state trajectory caused by uncertainty can be quantitatively represented separately. Next, analysis is performed based on the state transition tensor method. and The evolutionary pattern over time. Specifically, firstly, regarding... and Linearizing the state equations, we get: in, , as well as Let be the partial derivative with the corresponding subscript, i.e., the state transition tensor; similarly, linearizing the state deviation in the state update yields: Secondly, in order to conduct unified processing, and State augmentation is performed as follows: The state equation for the augmented state is: in, The covariance matrix of the state equation for the augmented state is: in, Let be the covariance matrix of the system state deviation. It is the identity matrix. It is a zero matrix. for and The covariance matrix of the augmented state vector; Finally, based on the covariance matrix of the system state deviation The state transition tensor is solved and covariance propagation is performed. The extreme values ​​of the system state deviation are calculated to obtain the final trajectory deviation range within the fault range, that is, to determine the future orbital reachability domain of the rocket.

[0128] During rocket flight, the mission area is divided into several different regions or phases according to different mission requirements and flight stages to facilitate effective trajectory planning, monitoring, and control. The division of each mission area is usually related to specific flight objectives, phases, or constraints, such as entering the target orbit, performing transition maneuvers, and controlling flight speed and altitude. However, when a rocket experiences a sudden malfunction, the original launch trajectory mission area faces the risk of failure. In this case, the launch trajectory mission area is no longer limited to the specific actions performed during rocket flight, but focuses more on whether the launch mission can be completed according to the established plan. Therefore, based on the rocket's health status and flight state after a sudden malfunction, the launch trajectory mission needs to be divided into main mission zones such as the target orbit, the degrade orbit, and the rescue orbit. To ensure the accuracy of these zones, extensive simulation experiments are needed to assess the orbital reachability envelope under abnormal rocket conditions, thereby improving online decision-making efficiency by building an offline knowledge base. The following section will introduce the process of calculating the launch vehicle's orbital reachability envelope through simulation experiments under conditions containing uncertain influencing factors: First, when a thrust reduction failure occurs, the rocket control system obtains an estimated thrust reduction ratio. , The estimated values ​​often contain significant uncertainties. Furthermore, to improve the prediction accuracy of the rocket's reachability domain, the uncertainty of the impact of a fault on thrust must be considered. Atmospheric density uncertainty and atmospheric pressure uncertainty Due to uncertainties, the true state of the launch vehicle's motion system is defined as follows: The equations of motion for a launch vehicle considering uncertainties are as follows: in, The velocity of a rocket relative to the atmosphere is expressed by the following formula: It is worth noting that, due to the nonlinear coupling of uncertainties in the rocket dynamics model, the weight matrix B of the uncertainties in the differential equations of the launch vehicle's dynamic system is replaced with a Jacobian matrix, thereby reducing the... and The linearized result of the state equation is updated as follows: in, Represents the Jacobian matrix; Similarly, the matrix w in the differential equation of the launch vehicle's dynamic system is also modified as follows: The thrust uncertainty and air density uncertainty satisfy a first-order Markov process, meaning the state at a given moment depends only on the state at the previous moment. in, , and These are the variances of the uncertainties in thrust, atmospheric density, and atmospheric pressure, respectively. and These correspond to the time constants of the deviation throughout the entire system.

[0129] In addition, the definition of the navigation state of a rocket's motion system includes the rocket's position, velocity, and atmospheric density, specifically: Navigation status Satisfying the dynamic equations: Based on this, the dynamic propagation equation of the state deviation over time and the covariance matrix of the system's actual state deviation can be obtained. In actual calculations, a problem of a large envelope boundary may be encountered, which can be addressed by adjusting the envelope by modifying the time constant.

[0130] Second, numerical integration can be used to determine the position covariance of the rocket terminal in the three axes of the coordinate system within the predicted time domain. Using the nominal orbit of the target mission as a reference orbit, and combining statistical methods... The criterion can be used to predict the reachability envelope of a rocket. Using the semi-major axis as the reference orbital trajectory as the center, several ellipsoids are used to represent the orbital envelope range of the launch vehicle.

[0131] By repeating the above steps, the reachable envelope range of the orbit can be obtained under multiple simulation experiments, and an offline knowledge base can be constructed based on this.

[0132] To address the feature extraction challenges in building offline knowledge bases, an intelligent classification method based on sparse stacked autoencoders (SSAEs) is proposed. Sparse stacked autoencoders (SSAEs), as a deep learning method, can effectively learn hierarchical representations of data through multi-layer nonlinear mapping, capturing complex patterns and high-order features in the input data.

[0133] The sparse stacked autoencoder uses an encoding module to process the input layer signal. Convert to hidden layer encoding Then, the decoding module converts it into output features. This enables the encoding process from the input layer to the hidden layer.

[0134] Sparse stacked autoencoder neural networks learn a... The function is one that attempts to approximate the output function to be close to the input function, and the mean squared error is often used as a loss function to describe this closeness: ; As can be seen from the loss function above, the standard sparse stacked autoencoder loss function uses mean squared error, which is not robust to data containing fault uncertainty. Therefore, this application first adds a certain sparsity constraint to the neurons in the hidden layer, enabling the autoencoder neural network to learn some representative features of the input signal more effectively. The sparsity constraint means that the output of most neurons in the network is close to 0, with only a small portion of neurons having effective outputs. This application uses the KL distance between the average activation and the set sparsity to achieve this sparsity constraint, which is specifically defined as follows: in, It is the weight that controls the sparsity penalty factor; Hidden layer neurons The average activation level, while the activation level of hidden layer neurons depends on... and ; This indicates the set sparsity. The closer the average activation is to the set sparsity, the smaller the penalty term for the KL distance. In addition, to avoid network overfitting, a weight decay term is introduced into the loss function to obtain smaller weights. This improves the generalization ability of the network model. Its specific definition is as follows: in, This is the weight decay term. yes Element; The coefficients for controlling the weight constraints; Indicates the first The number of nodes in the layer; It can also be expressed in the following form: in, This represents the loss function that does not include weight decay. yes Regularization term; Based on this, the loss function of the sparse stacked autoencoder can be obtained as follows: ; for In a layer-stacked autoencoder neural network, the encoding process of a sparsely stacked autoencoder neural network executes the encoding steps of each layer of the autoencoder in a front-to-back order: Similarly, the decoding process of a sparsely stacked autoencoder neural network executes the decoding steps of each autoencoder layer in a backward-forward order: in, It is the activation value of the deepest hidden unit, which contains the most valuable information in the data and constitutes a high-order feature representation of the input data. By... The features are used as input to the SoftMax classifier to obtain the classification result. Specifically, by inputting the aforementioned reachable domain data into a sparse stacked autoencoder neural network, reachable domain features are extracted, ultimately achieving task partitioning.

[0135] The training process of a sparse stacked autoencoder neural network mainly includes: First, unsupervised layer-by-layer pre-training is performed, starting from the first layer of the network. The hidden layer features learned by each layer are used as the input of the next layer. Layer-by-layer pre-training can obtain better initial parameters and better adapt to the structural features in the data, so that the entire network has a better initial state in the final supervised learning, thereby improving the iterative efficiency and stability of the algorithm. Then, each pre-trained hidden layer is connected, and the output of the final encoding layer is used as the input of the SoftMax layer. The output of SoftMax corresponds to the label of the data, and supervised fine-tuning training is performed. Finally, the parameters of the entire network are adjusted through the backpropagation algorithm.

[0136] ② Launch trajectory reconstruction module The launch trajectory reconstruction module is used to replan the launch trajectory based on the launch trajectory mission area, and reconstruct the optimal launch trajectory for the launch vehicle to complete the original predetermined orbital mission in the event of a sudden failure. For the rocket launch trajectory reconstruction problem after a sudden failure, this module first establishes a dynamic and constraint model that considers the impact of the failure, transforming the launch trajectory reconstruction problem into an optimal launch trajectory planning problem. Finally, to ensure the speed and convergence of the solution, the launch trajectory reconstruction problem is discretized and convexized by combining the local optimality characteristics of convex optimization, and a convex optimization method is used to achieve a fast online solution over multiple failure modes and long time domains.

[0137] First, for ease of description, a dynamic model of the launch vehicle is established in the launch coordinate system. In this model, the launch vehicle is treated as a point mass undergoing three degrees of freedom of translation. At any given moment, the position, velocity, and mass of the launch vehicle are represented by... , and The thrust of a launch vehicle is expressed in dynamics as the magnitude of the thrust. and the unit vector of the thrust direction The product of these two terms, in the launch coordinate system, the dynamic equations of the launch vehicle can usually be expressed as: in, This indicates the air resistance experienced by the rocket; The gravitational constant of Earth; This refers to the engine's specific impulse; Let be the average gravitational acceleration at sea level. The formula for calculating air resistance is: in, This is the drag coefficient; The area is the reference area for resistance. This is the Earth's rotational angular velocity; Expressed as atmospheric density, its calculation formula is: In the formula, This represents the atmospheric density at sea level. The altitude at which the rocket is located; This is the density reference height.

[0138] Secondly, during each segment of the launch vehicle's ascent phase, the rocket's mass should be constrained to be greater than its dry weight. That is, the mass constraint of the rocket is: Based on the changes in thrust, mass, and other parameters during different flight phases, the rocket's ascent trajectory is divided into several distinct stages. The changes in states at the transitions between these stages are also considered part of the system constraints. In the state variables connecting one stage to the next, only mass changes; other states, such as position, velocity, and thrust direction, remain consistent. Therefore, the constraints at the stage transitions can be characterized as follows: The superscript of each variable Indicates the stage number; This marks the beginning of the current stage. This refers to the final moment of the current stage. This represents the amount of quality loss at the final moment of the current stage. In the event of a thrust failure, the top priority is to ensure the rocket still reaches its intended mission orbit. Therefore, constraints on the orbital insertion position must be established at the end of the trajectory. The terminal orbital insertion constraint of the rocket's ascent trajectory can be expressed as: in, and The vectors representing the angular momentum and eccentricity of the orbital path entering the orbit from the original nominal trajectory; Furthermore, the coplanar constraint of the orbits is expressed as: For the rocket thrust direction vector If the magnitude of the vector should be equal to 1, then its constraint can be expressed as: ; When faced with different thrust failure scenarios, launch vehicles need to select different estimation and optimization strategies based on the severity of the failure. Therefore, trajectory optimization after a thrust failure can be categorized into two main types based on different failure scenarios and optimization objectives: i) Using fuel efficiency as the optimization objective function ii) Orbital insertion rescue track, with track height as the optimization objective function. To ensure the efficiency and convergence of solving the optimal launch trajectory planning problem, this application adopts a first-order discretization-preserving method to transform the above nonlinear nonconvex problem into a convex problem.

[0139] use Representing the discrete points within each stage, the discretized dynamic constraints can be expressed as: Under dynamic constraints, air drag, atmospheric density, and the discrete form of the rocket's altitude can be characterized as follows: The quality constraints in each stage are discretized as follows: The discrete form of the state constraints at the connection between each adjacent stage is: Furthermore, the terminal orbit insertion constraint and the thrust direction unit vector constraint are discretized and expressed as: ; The discretized dynamic equations and constraints still exhibit strong nonlinearity. To improve the solution efficiency and convergence of the optimization problem, it is necessary to transform the trajectory replanning problem into a linear convex problem as much as possible. Therefore, a sequential convexization approach is used for the transformation. Furthermore, while linearizing the optimization problem allows it to be solved within the convex optimization premise, it also introduces two significant problems: unbounded solutions and the appearance of pseudo-feasible solutions. To address these issues, two solutions—confidence region constraints and variable constraint relaxation—are proposed to improve the accuracy and convergence of the solution.

[0140] To address the potential unboundedness of the solution results caused by local approximation methods used in the linearization process, trust region constraints are introduced to limit the search area. This ensures that the solution obtained in each iteration does not deviate too far from the current solution, avoiding inaccurate solutions due to linearization errors. Specifically, the location of the rocket trajectory... and speed Essentially, the direction vector of the thrust and the time interval between stages Due to the influence of [the pandemic], the confidence region constraint is only set in [the following context]. and Therefore, the confidence region constraint can be expressed as: Wherein, the subscript of the variable Indicates the number of iterations; and The purpose of these constraints is to limit the differences between the corresponding variables to make them bounded. Therefore, in subsequent iterative optimization solutions, these constraints will eventually be extended to the objective function.

[0141] Furthermore, to prevent spurious solutions from arising during linearization, especially in the nonlinear components, adaptive slack variables are added to the system. The physical meaning of this slack variable can be understood as the compensation acceleration or virtual control quantity of the system when the calculated thrust feasibility of the launch vehicle is insufficient. The confidence constraints of this variable and the calculation formula for its corresponding variable are expressed as follows: To improve the convergence and feasibility of the computation process, relaxation parameters are also added to other state constraints. The terminal trajectory constraint after relaxation and convexity treatment can be expressed as: in, and These represent the solution accuracy in the two constraints mentioned above, respectively. The thrust direction vector constraint can be relaxed as follows: ; For the trajectory optimization problem of a multi-stage launch vehicle, the initialization calculations before the actual computation are based on the nominal trajectory under fault-free conditions to obtain the state values ​​required for the optimization solution; the most important initial state variables include the rocket position. ,speed and thrust direction unit vector The core idea of ​​initialization is to ensure that the trajectory in the calculation process satisfies the dynamics and constraints as much as possible under the compensation of relaxation terms. The addition of initialization increases the convergence of the algorithm.

[0142] Furthermore, the thrust magnitude in this optimization model is not a constant value, but is related to the atmospheric pressure at the location. In order to satisfy the convex constraints while ensuring the feasibility of the solution, the atmospheric pressure term is handled by an iterative approximation method. That is, in each optimization calculation, the atmospheric pressure value at the initial optimization value (the optimization result of the previous iteration) is taken as the numerical input for this time, and then approximation is performed through iteration.

[0143] Based on the above descriptions of the fault trajectory optimization problem, the final discretized convex trajectory optimization model is as follows: Objective function: Constraints: 1) Initialize constraints: 2) Dynamic constraints: 3) Constraints at the connection between adjacent stages: 4) Terminal orbital insertion constraints: 5) Thrust direction vector constraint: 6) Confidence region constraint: .

[0144] Finally, by solving the above discretized convex trajectory optimization model using the convex optimization method, the optimal trajectory for the launch vehicle to complete its original intended orbital insertion mission under the condition of a sudden failure can be obtained.

[0145] ③ Control Reconfiguration Module The control reconfiguration module is used to reconfigure the flight trajectory control law and attitude control law, thereby controlling the launch vehicle's flight attitude and trajectory in the event of a sudden failure.

[0146] In this system, the input information for the control law reconstruction module includes: ①Flight status data: real-time attitude, angular velocity, orbital parameters, etc.; ② Actuator status information: Availability, performance degradation, and health status assessment results of actuators such as thrusters and reaction flywheels; ③ Tasks and constraints: attitude control accuracy requirements, maneuver time window, fuel and power consumption limits, etc.; ④ Fault detection results: Fault type, occurrence time, and degree of impact.

[0147] The flight status data, actuator status information, tasks and constraints are derived from the rapid decision-making system, and the fault detection results are derived from the intelligent diagnostic system.

[0148] After being fused and preprocessed, the input data drives the control law reconstruction module to calculate new control allocation schemes in real time, so as to maintain the spacecraft's attitude control performance and mission execution capability under the conditions of actuator failure or performance degradation.

[0149] During launch, rockets require precise control of their flight trajectory to ensure successful atmospheric entry and entry into their designated orbit. Due to the complex flight environment, the variability of rocket systems, and the high probability of malfunctions, rocket control systems typically incorporate dynamic redundancy mechanisms. These mechanisms ensure that if a component of the rocket system fails (such as the attitude control system, thrust control system, or sensors), and the original control strategy becomes ineffective, the system can adjust the flight path through control reconfiguration, ensuring the rocket can still complete its mission even under fault conditions.

[0150] Traditional control reconfiguration methods, such as PID control or model-based control methods, typically rely on accurate system models and fixed control gains to stabilize flight trajectories. This results in a lack of sufficient flexibility and robustness, limited fault tolerance for system failures, and difficulty in coping with complex faults and environmental changes that may occur during rocket flight. This application employs an adaptive augmentation method for control parameters to achieve control law reconfiguration.

[0151] When an inertial navigation system (INS) malfunctions during rocket flight, attitude angle information cannot be directly obtained. Therefore, an attitude angle observer is designed to estimate angular velocity and reconstruct the attitude control law. Considering the high-order strong nonlinear characteristics of the rocket model, a complementary filter observer is used to estimate the angular velocity. This observer utilizes data from the accelerometer and magnetometer, combined with the control input, to estimate the attitude angle and angular velocity.

[0152] In the event of a control system failure, adaptive augmenting control (AAC) is employed to reconstruct the attitude control law. AAC combines the advantages of adaptive and augmented control, enabling adjustments to the control strategy based on real-time flight conditions when a system failure occurs. When a launch vehicle experiences a sudden malfunction, AAC can estimate the flight state and fault severity in real time using fault diagnosis results. Combined with the mission decision scheme from the launch trajectory reconfiguration module and the launch trajectory replanning results, it dynamically adjusts the control gain and control strategy, ensuring sufficient robustness of the control system. Furthermore, through adaptive gain adjustment, AAC can adjust the control gain in real time after a failure, ensuring the system quickly recovers to its optimal state.

[0153] The baseline controller of the adaptive augmented control (AAC) system is a PD+correction network, and its design method is the same as that of traditional rocket control controllers. The main function of the baseline controller is to ensure the stability of the rocket attitude control system under nominal conditions and to meet certain frequency and time domain performance indicators. The reference model in the AAC control framework is a standard second-order system used to simulate the dynamic process of the rocket's launch trajectory response under ideal conditions. Subtracting the actual trajectory response of the rocket from the output of the reference model effectively removes rigid body motion signals from the feedback signal, avoiding the influence of rigid body motion signals on the signal extraction process.

[0154] It is worth noting that the adaptive gain law is the core of adaptive augmentation control. It adjusts the forward gain of the controller based on the signals output by the high-pass and low-pass filters. When the error signal increases, the adaptive gain law increases the forward gain, raising the system crossover frequency and improving the system response speed to reduce tracking error. When the rocket body vibration signal intensifies, the adaptive gain law decreases the forward gain, lowering the system crossover frequency and frequency domain response amplitude, reducing the excitation on elastic vibration and thus suppressing rocket body vibration. When the error signal is small and the vibration intensity is low, the adaptive law adjusts the forward gain coefficient to 1 to reduce its impact on the control system. Furthermore, considering that excessively high or low forward gain can affect system stability during gain adjustment, upper and lower bounds for gain adjustment and a gain limiting function are introduced. When the gain approaches the gain adjustment boundary, the limiting function restricts the direction of gain change, ensuring that the forward gain does not exceed the upper and lower bounds.

[0155] When a system fault is detected, the rocket control system will reconstruct the attitude control law based on an adaptive augmented control method. The core of adaptive control is gain adjustment, that is, dynamically adjusting the control gain to cope with changes in the system state. In rocket control, a PID controller is used for basic control, and the gain matrix is ​​adaptively adjusted. , and This allows for dynamic adjustments.

[0156] The standard PID form of the control variable is: in, This indicates the error in the rocket's deviation from its nominal trajectory.

[0157] In rocket control systems, reference models are typically used to provide the desired trajectory or flight state. For example, if the rocket's mission is to enter a specific orbit, the reference model generates the desired trajectory based on a predetermined flight path. The AAC (Autonomous Launch Control) system compares the actual flight trajectory with the desired trajectory from the reference model to determine errors in the flight state and adjusts the control strategy in real time. To optimize the launch trajectory, the optimal trajectory output by the launch trajectory reconstruction module is used as the reference model.

[0158] Furthermore, based on rocket dynamics and reference model information, an adaptive gain regulator is designed to enable the system to dynamically adjust the controller gain according to the target behavior provided by the reference model, thereby ensuring optimal system response under different flight conditions. When the system malfunctions or is disturbed, the deviation between the "ideal" state provided by the reference model and the actual state is fed back in real time, guiding the direction of gain adjustment. When the error is large, adaptive control can increase the control gain to improve the system response speed; while when the system recovers to a steady state, the gain is appropriately reduced to prevent over-control. Commonly used error-based gain adjustment methods are as follows: in, , , These are the initial values ​​for the proportional, integral, and differential gains, respectively. , , These are respectively related to error Related gain adjustment amount.

[0159] Gain adaptive laws can adjust the system's forward gain, improving the controller's adaptability to internal and external disturbances. However, if the gain adjustment exceeds a certain level, it can lead to system instability. Therefore, designing upper and lower bounds for gain adjustment ensures that the gain remains within a suitable range, guaranteeing controller stability and providing sufficient margin.

[0160] ① Increasing the gain will lead to a decrease in the gain margin and phase margin of the rigid body; the phase margin of phase-stable systems will decrease; and the gain margin of amplitude-stable systems will decrease. Therefore, to meet the general requirements of the control system, the upper bound of the gain is... When selecting, the phase margins should generally be considered. Each amplitude margin .

[0161] ② When the gain decreases, the rigid body cutoff frequency decreases; the gain margin and low-frequency gain margin of the rigid body also decrease. Therefore, the lower bound of the gain... When selecting, the rigid body shear frequency should generally be considered. Phase margin .

[0162] After determining the upper and lower bounds of the gain, to ensure the stability and response performance of the control system after gain adjustment, a gain limiting function needs to be designed. The gain limiting function must satisfy the following conditions: 1) the gain is close to 1 within the upper and lower bounds; 2) the gain rapidly decreases to 0 when approaching the boundary; 3) the function is smooth. Therefore, consider a family of functions with a "cliff" property, which has the following form: in, , As a parameter, and .

[0163] To facilitate adjusting the curve characteristics of the function, the above equation can be transformed into: For the transformed function, the parameters The parameter determines the steepness of the function; It allows functions to be shifted left or right.

[0164] Furthermore, the design requirements for the upper and lower bound functions of gain are transformed into "cliff" function parameters. Suppose that the upper bound function of gain needs to... Within the range, from to ,but: achievable Similarly, suppose the lower bound function of the gain needs to be... Within the range, from Rise to ,but Express the upper and lower bound functions of gain in a unified form in, or , , Note that, although in the formula The item can be simplified to However, considering Since it is a constant, most microcontrollers or processors have trigonometric arithmetic unit modules that can speed up the calculation of the exp function, so the expression is not further simplified.

[0165] Since its inception, AAC (Adaptive Gain Control) has had two typical forms of gain adaptive law. This application adopts the second-generation gain adaptive law to implement the control law design. The second-generation gain adaptive law is as follows: in, , and It is a constant parameter.

[0166] The second-generation adaptive law uses a gain limiting function, which makes the gain boosting and gain debuffing terms symmetrical and simpler. It ensures both the separation of variables over a large range, facilitating analysis, and that the gain adjustment is within the limits.

[0167] Based on the above analysis, and combined with the rocket's dynamic model and adaptive augmented control method, the control system can adjust the control gain and thrust in real time when a sudden failure occurs, ensuring that the rocket can continue to fly stably.

[0168] Furthermore, when the thrust of a rocket engine decreases or even fails, the thrust allocation scheme needs to be redistributed. With the goal of minimizing fuel consumption during the redistribution of the yaw angle, the yaw angle control redistribution problem can be transformed into a performance-constrained optimization problem as shown below: in The engine's swashplate angle, as can be determined from the control distribution method: in, U To calculate the control quantities required for each attitude channel in the actual control system, K For the allocation matrix.

[0169] Under the condition of satisfying dynamic constraints, the above performance index constraint optimization problem is solved by using the improved simplex method to obtain the optimal thrust distribution scheme.

[0170] During rocket attitude control, the core stage and booster engines distribute the total pitch angle command generated by the controller in a certain proportion. Therefore, to further simplify the model, the total equivalent pitch angle of the pitch channel is introduced as a control variable. Its relationship with the equivalent pitch angle generated by the core stage and booster engines is as follows: The single-input fault dynamics model is obtained: in, This indicates the additional disturbance force and disturbance torque caused by the additional swing angle.

[0171] It can be seen that under this fault modeling method, the input-output structure of the control system does not change before and after the fault; only the external disturbance term caused by the fault is added to the outside of the system.

[0172] In the single-input fault dynamics model, the first equation represents the normal motion of the rocket's center of mass, and the second equation represents the motion of the pitch channel around the center of mass. During rocket flight, the changes in various motion parameters with disturbances exhibit both long-period and short-period variations, the different nature of which is determined by their inherent characteristics. This is because, after the rocket is disturbed, the changes in angular motion parameters and center of mass motion parameters occur in two different ways: the angular parameters caused by the disturbance change very rapidly, while the velocity changes very slowly. Therefore, in designing the adaptive controller parameters, the motion of the rocket's center of mass can be ignored initially, considering only the motion of the rocket around the center of mass. Thus, without controlling the external disturbance term, and using the total equivalent pitch angle of the pitch channel as the control variable, the model of the second-order control system is finally obtained: Similarly, in the scrolling channel, let: The fault dynamics model of the rolling channel after the fault can be obtained: in, This indicates the additional disturbance torque generated in the rolling channel.

[0173] It should be noted that although a single servo mechanism jamming will cause the control of the two channels to be coupled, such as the jamming of the No. 2 booster engine affecting the attitude control of both the pitch and roll channels, the control structures of the two channels are not coupled under the fault modeling method given above. Only disturbance terms are added to the two channels respectively. Therefore, adaptive controllers can be designed for the two channels separately.

[0174] Therefore, adaptive controllers were designed for the three channels respectively. When a fault occurs, the controller can adaptively adjust the controller parameters online to eliminate the effects of additional interference.

[0175] Let the state equation of the system be: in, for 3D state vector; for A constant matrix. Choose the following quadratic form functions as possible Lyapunov functions: in, for Symmetric positive definite matrix, function The derivative with respect to time: in, Therefore, a sufficient condition for a linearly stable system to be stable is Q It must be a positive definite matrix. Usually, a positive definite matrix is ​​chosen first. Q Array, solve P And then according to P Whether a system is positive definite is used to determine its asymptotic stability. For ease of calculation, Q The array is usually taken as an identity array or a diagonal array.

[0176] Based on the Lyapunov stability criterion described above, an adaptive fault-tolerant controller is designed, taking the control of the pitch channel as an example. The design methods for the other two channels are similar.

[0177] The adjustable part is controlled by a combination of feedforward and feedback control. The feedforward controller uses proportional gain, and the feedback controller uses a PD controller.

[0178] The reference model is a stable system with ideal control performance, and it is generally designed using the pole placement method. The open-loop transfer function of the second-order system can be easily obtained as follows: To better observe the fault-tolerant performance of the control system, we use the design method of the pitch channel controller under normal conditions as the parameters for pole placement, that is, using attitude angle deviation and attitude angular velocity feedback control: The transfer function of the parametric model is obtained as follows: The corresponding differential equation is in, Because the dynamic coefficients of a rocket vary significantly during the boost phase, the steady reference model typically used in traditional adaptive controller design is often insufficient to meet system performance requirements, and the coefficients of the reference model are still set to time-varying coefficients. This necessitates that the control system calculates the reference model online at each step and then adjusts the controller parameters accordingly.

[0179] For the aforementioned adjustable system, the following control structure is adopted: in, For feedforward gain; , For feedback gain, parameters are adaptively adjusted. , and Make the output of the controlled object track the output of the reference model.

[0180] The differential equations of the adjustable system are obtained. Define generalized error The systematic error equation is obtained as follows: in, Introducing parameters of the generalized error vector machine The generalized error equation can be written in matrix form: Choose the Lyapunov function as in, It is a two-dimensional positive definite matrix; It is a three-dimensional positive definite matrix, and its specific expression is: The derivative of the Lyapunov function is: In order to Choose a positive definite matrix satisfy At the same time, the adaptive controller is selected as: but With a negative constant, the closed-loop system is asymptotically stable.

[0181] Compared to changes in the rocket's flight attitude angle, the coefficients in the dynamic equations change very slowly, that is... , , Therefore, it can be considered that... , , Thus, the feedback gain is obtained. , and feedforward gain coefficient The adaptive law is: For the adaptive control law in the differential form described above, integration yields the expression for the controller parameters. Similarly, the adaptive control law for the yaw channel can be derived as follows: The adaptive control law for the rolling channel is The above-described scheme can adjust controller parameters in real time and dynamically adapt to changes in system state, effectively addressing various faults and uncertainties in the rocket attitude control system. Compared with traditional methods, it has stronger robustness and flexibility, can handle multiple fault modes, and ensures that the rocket maintains a stable attitude during flight to achieve mission objectives.

[0182] Furthermore, the reconstructed attitude control law controls the flight trajectory based on the optimal trajectory replanned under sudden failure, thereby realizing the reconstruction of the flight trajectory control.

[0183] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0184] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0185] The above provides a detailed description of the experimental system for online autonomous monitoring and intelligent planning of aerospace launch vehicles provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An experimental system for online autonomous monitoring and intelligent planning of space launch vehicles, characterized in that, The system comprises a user layer, a business layer, and a data access layer, wherein: The user layer is used to provide users with a graphical user interface. The business layer is used to receive and process business requests from the user layer, and to perform data calculations, analysis, and task planning based on the business requests. The data access layer is used to manage local databases or data files and handle the reading, writing, and accessing of related data.

2. The system according to claim 1, characterized in that, The system also includes hardware components, wherein the hardware components include: The onboard computer is used to perform various flight control, mission management, and fault diagnosis tasks. Equivalent simulator, used to simulate spacecraft flight status and fault scenarios; Functional boards are used to implement data acquisition, control execution, and signal processing; Sensors and actuators collect flight data from the launch vehicle through different sensors and control the attitude adjustment of the spacecraft through actuators.

3. The system according to claim 2, characterized in that, The business layer includes: The spacecraft simulation and fault data generation system is used to simulate the flight process of launch vehicles and can simulate typical faults that rockets may encounter during flight. The online intelligent diagnostic system is used to monitor the onboard status and diagnose online faults based on key data output by the system generated from spacecraft simulation and fault data. The online intelligent planning system is used to re-divide the launch trajectory mission area and re-plan the launch trajectory and control law when a launch vehicle malfunctions, and adjust the launch vehicle's flight attitude and orbit.

4. The system according to claim 3, characterized in that, The spacecraft simulation and fault data generation system includes: The fully digital simulation module for the flight process is used to simulate the six-degree-of-freedom ballistic motion, dynamic characteristics, aerodynamic characteristics, and external environment of a spacecraft, and obtain the flight status data of the spacecraft. The sensor data generation module is used to generate key sensor simulation data based on the spacecraft's flight status data and key sensor models; The typical fault modeling module is used to model potential fault conditions based on simulation data from key sensors.

5. The system according to claim 4, characterized in that, The online intelligent diagnostic system includes: The autonomous health assessment module is used to preprocess the acquired flight status data and sensor simulation data of the spacecraft, and to monitor the status based on the preprocessed standardized data, and to issue alarms for any abnormal states. An autonomous fault identification module is used to analyze and identify received abnormal state alarm information to obtain fault classification results; the fault classification results include fault type, location and severity. The fault handling decision module determines the fault index vector based on the fault classification results, and maps the fault index vector into fault handling instructions through a decision mapping function.

6. The system according to claim 5, characterized in that, The preprocessing of the acquired sensor data includes: For raw sensor data with uneven distribution, the Lagrange polynomial interpolation algorithm is used to fill in the missing parameters; An improved wavelet threshold function is constructed by introducing exponential functions with different bases. The incomplete rocket sensor sample data is denoised based on the improved wavelet threshold function.

7. The system according to claim 6, characterized in that, The improved threshold function is: in, As a regulating factor, T For the threshold, To improve the threshold function after processing the first... The first in scale Wavelet coefficients, To improve the threshold function before processing the first... The first in scale Wavelet coefficients.

8. The system according to claim 6, characterized in that, The decision mapping function is: Among them, each parameter matrix With threshold All data were obtained through offline data analysis and online adaptive correction. To provide comprehensive fault indicators, This is the baseline value vector for each indicator.

9. The system according to claim 7, characterized in that, The online intelligent planning module includes: The launch trajectory mission reconfiguration module is used to determine the reachability domain of the launch vehicle in an abnormal state when the launch vehicle malfunctions, and to repartition the mission based on the reachability domain data. The launch trajectory reconstruction module is used to replan the launch trajectory based on the mission area and mission partitioning, and reconstruct the optimal launch trajectory for the launch vehicle to complete the original predetermined orbital mission in the event of a sudden failure. The control reconfiguration module is used to reconfigure the flight trajectory control law and attitude control law, thereby controlling the launch vehicle's flight attitude and trajectory in the event of a sudden failure.

Images