Space debris on-orbit real-time autonomous avoidance method, system and spacecraft

Abstract

The application discloses a kind of space debris on-orbit real-time autonomous avoidance method, system and spacecraft, belong to spacecraft on-orbit safety control technical field.The method is solidified and burned in on-board chip by light-weight AI model of knowledge distillation, structured pruning and 8-bit quantization;Adopt "standard strategy library+AI fine-tuning" architecture, call pre-stored orbit lag / advance / out-of-plane maneuver template, generate candidate trajectory by on-orbit fine-tuning of template parameter by solidified AI;Safety check is carried out based on simplified dynamics model, to ensure that meet off-target distance, fuel and attitude stability constraint;Build double-threshold emergency redundancy mechanism, select fuel optimal trajectory generation instruction execution in normal mode, when collision probability is too high or chip is abnormal, immediately call pre-stored emergency instruction sequence, directly execute after being quickly analyzed by propulsion controller.The application shortens response time from hour level to minute level, significantly improves the autonomous survival ability and mission reliability of spacecraft under extreme working conditions.

Description

Technical Field

[0001] This invention relates to the field of spacecraft on-orbit safety control technology, specifically to a method, system, and spacecraft carrying the system for real-time autonomous avoidance of space debris in orbit. Background Technology

[0002] With the rapid increase in the number of spacecraft in low Earth orbit, the space debris environment is deteriorating, posing a serious threat to the operational safety of spacecraft in orbit. Addressing the risk of space debris collisions has become an indispensable and critical aspect of space mission design and on-orbit operation management.

[0003] (1) Traditional ground-based decision-making model and its shortcomings:

[0004] Currently, the vast majority of spacecraft still adopt the "ground-based decision-making" model, with the ground mission control center at its core. A typical process includes: (a) debris collision warnings are provided by the ground-based space surveillance network; (b) the warning data is sent to the ground mission control center; (c) ground personnel use high-performance computing resources to perform several hours of precise orbit determination, collision probability calculations, and avoidance strategy simulations; (d) after the strategy is determined, the spacecraft waits for the telemetry, tracking, and command (TT&C) segment to transmit the commands to the spacecraft; and (e) the spacecraft executes the commands.

[0005] This traditional model has inherent flaws:

[0006] Slow response: Due to the large number of manual decision-making processes, cross-system coordination, and waiting for monitoring and control windows, the entire process response time can take several hours or even more than ten hours, making it unable to cope with sudden, close-range debris collision threats.

[0007] Weak survivability: The strong dependence on ground-based telemetry and control systems and space-to-ground links makes spacecraft extremely vulnerable in extreme situations such as wartime, electronic warfare, or interruption of space-to-ground communication, resulting in insufficient autonomy and survivability.

[0008] High resource consumption: It is difficult to efficiently cope with the massive and high-frequency collision avoidance requirements faced by a large-scale constellation of tens of thousands of satellites in the future, and the ground system operation is under heavy burden.

[0009] (2) Shortcomings of existing on-board intelligent avoidance technology:

[0010] To enhance autonomy, existing technologies have included research on applying artificial intelligence to collision avoidance decision-making. However, these solutions still have significant limitations, which can be summarized as "having algorithms but no practical application" or "having intelligence but not practical application".

[0011] High computational complexity makes on-board deployment difficult: For example, the deep reinforcement learning models such as PPO and LSTM used in Document 1 (CN116125811A) and Document 2 (CN117972901A), while advanced in their algorithms, are structurally complex and have huge computational and storage requirements. They are essentially computationally intensive tasks and cannot meet real-time requirements on resource-constrained on-board computing platforms. These solutions mostly focus on improving the algorithm simulation level and lack a comprehensive engineering consideration of on-board edge computing hardware, extreme model lightweighting, and on-board hardware-software collaboration.

[0012] True single-satellite autonomy not achieved: Another technical approach, such as document 3 (CN120651250A), attempts to improve real-time performance by building a distributed satellite edge computing network. While this approach reduces reliance on ground systems to some extent, its decision-making center is still distributed within the constellation network, relying on continuous inter-satellite communication, and thus does not achieve true single-satellite autonomous closed-loop operation. In extreme cases such as inter-satellite link interruption, the system's collision avoidance capability will decrease sharply.

[0013] While existing technologies (such as CN120364159A) propose the concept of onboard AI decision-making, they still suffer from the following key shortcomings: First, lack of engineering feasibility: Existing solutions only generalize to "deployment on onboard computers," without disclosing how to adapt complex neural network models to aerospace-grade FPGA or ASIC hardware with extremely limited computing resources, storage space, and power consumption. Directly porting uncompressed models will lead to excessive inference latency or hardware resource overflow. Second, lack of physical interpretability and reliability in the decision-making process: Existing end-to-end AI decision-making models are pure "black boxes," and their output avoidance trajectories may violate basic orbital mechanics principles. Third, lack of system-level safety redundancy: Existing solutions rely on a single AI decision-making link and do not consider hardware failures such as single-event upsets and logical disorder that may occur in the deep space environment, resulting in a weakness in system survivability.

[0014] Therefore, there is an urgent need in this field for a new method and system that can truly adapt to the limited computing, storage, and power consumption resources onboard, and achieve fast, reliable, and low-power onboard autonomous obstacle avoidance. This method not only needs to have intelligent decision-making capabilities, but more importantly, it must be feasible for onboard engineering, that is, "having both the algorithm and the platform". Summary of the Invention

[0015] This invention aims to solve the above problems and provide a space debris autonomous avoidance scheme with engineering feasibility, physical interpretability, and system-level redundancy.

[0016] To achieve the above objectives, the present invention provides the following technical solution:

[0017] A method for real-time autonomous avoidance of space debris in orbit includes the following steps:

[0018] S1, Information Acquisition and Selection of Different Avoidance Procedures Based on Real-Time Collision Probability Estimation: Receive space debris collision warning information from at least one information source, acquire the spacecraft's own status information, and input it into the onboard hardware acceleration computing platform to estimate the collision probability in real time.

[0019] S11, Normal Mode: When the estimated collision probability is higher than the first threshold, execute the normal optimization avoidance process from step S2 to step S6.

[0020] S12, Emergency Mode: When the estimated collision probability is higher than the second threshold, the second threshold is greater than the first threshold, or an abnormal operation of the onboard hardware acceleration computing platform is detected, the emergency response process is executed: skip the optimization calculation of the lightweight AI model deployed on the onboard hardware acceleration computing platform, and directly call and execute the pre-stored emergency evasion maneuver instruction sequence from the onboard memory.

[0021] S2, Deployment of lightweight AI model with hardware: Input the warning information and the self-state information into the lightweight AI model;

[0022] S3, Hybrid Intelligent Decision-Making: Generates candidate trajectories for evasion maneuvers;

[0023] S31, call the basic avoidance strategy from the standard avoidance strategy library pre-stored on the satellite. The standard avoidance strategy library contains at least the orbital lag maneuver, orbital advance maneuver, and out-of-plane maneuver templates constructed based on prior knowledge of orbital mechanics.

[0024] S32, the warning information and the self-state information are input into the lightweight AI model embedded in the on-board hardware acceleration computing platform. The model performs on-orbit real-time fine-tuning of the maneuver parameters of the basic evasion strategy template to generate at least one physically feasible evasion maneuver candidate trajectory.

[0025] S4, Fast Security Verification:

[0026] Based on a simplified orbital dynamics model, millisecond-level fast forward simulation and safety verification are performed on the at least one candidate trajectory for evasion maneuvers. The constraints for the safety verification include at least the following:

[0027] The minimum miss distance from the debris after avoidance is greater than the preset safety threshold;

[0028] The total velocity increment required for the candidate trajectory is less than the spacecraft's current remaining fuel equivalent velocity increment;

[0029] The estimated attitude maneuver angle and angular velocity during the maneuver are within the stable control range of the onboard attitude control system;

[0030] S5, Optimal trajectory selection: Select the trajectory with optimal fuel consumption from the candidate trajectories after safety verification;

[0031] S6, the pulse maneuver control command is directly sent to the on-board propulsion system for execution to complete the autonomous evasive maneuver.

[0032] Preferably, step S2 includes:

[0033] S21, in the ground phase, compresses the complex teacher model into a lightweight student model with less than 1M parameters and less than 4MB in size through knowledge distillation, structured pruning and 8-bit quantization techniques.

[0034] S22, the lightweight student model is burned into the logic unit of the onboard FPGA or ASIC chip, making it a physical circuit capable of hardware-accelerated inference.

[0035] This invention also discloses a space debris real-time autonomous avoidance system for implementing the above method, comprising:

[0036] The onboard communication module is used to receive early warning information about multi-source collisions with space debris;

[0037] The onboard processing module is electrically connected to the onboard communication module, and its core processing unit is an FPGA or ASIC chip.

[0038] The FPGA or ASIC chip contains a lightweight AI model with fewer than 1M parameters and a size of less than 4MB, and stores a standard avoidance strategy library and an emergency avoidance instruction sequence.

[0039] The onboard processing module is configured to independently complete the entire decision-making process from threat assessment to command generation without ground intervention, and has dual-mode operation capability:

[0040] In the first mode, the lightweight AI model is run, and a conventional optimization and avoidance process is executed.

[0041] In the second mode, when the emergency response process triggering conditions are met, the lightweight AI model is bypassed, and the emergency evasion instruction sequence is directly invoked and executed.

[0042] The onboard processing module also includes a model update interface, which is used to receive model update data packets transmitted via the ground telemetry and control link, and to reconfigure some parameters and upgrade the version of the lightweight AI model embedded on the FPGA or ASIC chip.

[0043] The onboard execution module, electrically connected to the onboard processing module, includes a propulsion controller and a thruster, and is used to receive and execute maneuver control commands.

[0044] Preferably, the onboard processing module further includes a model update interface for receiving model update data packets transmitted via the ground telemetry and control link, and for replacing parameters or upgrading versions of existing lightweight AI models on the satellite.

[0045] Preferably, after completing the evasive maneuver, the system initiates a self-assessment of its safety status, using an onboard GNSS receiver or inter-satellite link to measure the actual trajectory after the evasion and compare it with the expected safe trajectory to confirm the risk of avoiding collision.

[0046] The present invention also discloses a spacecraft equipped with the aforementioned autonomous space debris avoidance system.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] First, it breaks through the barriers to engineering feasibility: through knowledge distillation, structured pruning and 8-bit quantization technology, the AI ​​model is compressed to less than 4MB and for the first time is solidified and burned into the logic unit of an aerospace-grade FPGA / ASIC chip, realizing on-board hardware accelerated inference with milliwatt-level power consumption and millisecond-level response under extreme resource constraints, and completely solving the industrialization bottleneck of "having algorithms but no carrier".

[0049] Second, it endows the decision-making process with physical interpretability and high reliability: The original "standard strategy library + AI fine-tuning" hybrid decision-making architecture ensures that any output of AI is rooted in a mature orbital mechanics template. AI only acts as a fine-tuning regulator to seek optimization within the physical feasible boundary, fundamentally eliminating the "outrageous" trajectory that pure black box AI may produce, and making the autonomous avoidance result both intelligent and engineering credible.

[0050] Third, a system-level dual safety redundancy is constructed: through a dual-threshold emergency redundancy mechanism of "AI intelligent optimization + hard-coded emergency bypass" and the introduction of hardware operation status monitoring, an independent hardware-level safety circuit breaker channel is provided for the main AI system, independent of software logic. When the main AI system fails for any reason (software anomaly, hardware failure, extreme threat), the spacecraft still has a deterministic minimum risk avoidance capability, fully meeting the highest level of functional safety and failure safety design specifications for space missions. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the autonomous evasion method of the present invention;

[0053] Figure 2 This is a diagram showing the core hardware components and data interaction relationships of the system of this invention;

[0054] Figure 3 This is a lifecycle diagram of the lightweight AI model of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The purpose of this invention is to provide a method, system, and spacecraft for real-time autonomous on-orbit space debris avoidance with hardware and software collaboration, oriented towards aerospace-grade FPGA / ASIC resource constraints, in order to solve the following technical problems:

[0057] Overcoming the engineering implementation gap: Solving the engineering challenge that existing intelligent collision avoidance algorithm models are too large and computationally complex to run in real time on resource-constrained and power-intensive aerospace-grade FPGA / ASIC hardware, by using extreme model compression and hardware solidification technology, advanced AI algorithms are made feasible for space engineering for the first time.

[0058] Imparting physical interpretability to decisions: Addressing the safety risks of uncontrollable end-to-end black-box AI decision-making results that may violate the laws of orbital mechanics, a hybrid architecture of "policy library guidance + AI fine-tuning" is used to ensure that AI output is always rooted in mature prior knowledge of orbital mechanics.

[0059] Constructing system-level safety redundancy: To address the risk of failure of a single AI decision-making link in the deep space environment due to hardware failures such as single-event upsets, a dual-threshold emergency redundancy mechanism is used to provide the main AI system with a hardware-level safety circuit breaker channel independent of software logic, significantly improving the spacecraft's autonomous survivability under extreme conditions.

[0060] 1. Overall technical architecture of the present invention:

[0061] The core of this invention lies in constructing an on-board autonomous avoidance system characterized by "hardware-based AI-driven architecture, policy library-guided safeguards, and dual-threshold emergency redundancy." Upon receiving an early warning, this system completes the entire decision-making chain, from threat analysis to maneuver execution, entirely on-board, without ground intervention. Its technical architecture comprises three innovative layers:

[0062] The bottom layer (hardware support layer) uses aerospace-grade FPGA or ASIC chips as the core, and the ultra-compressed lightweight AI model is solidified and burned into the chip logic unit to achieve hardware-accelerated inference with milliwatt-level power consumption and millisecond-level response.

[0063] The middle layer (hybrid decision-making layer) adopts a hybrid intelligent decision-making architecture of "standard strategy library + AI fine-tuning". The strategy library provides basic maneuver templates based on prior knowledge of orbital mechanics, and the AI ​​model only fine-tunes the template parameters to ensure that the decision results are both intelligent and physically feasible.

[0064] Top layer (safety redundancy layer): Construct a dual-threshold emergency redundancy mechanism, monitor collision probability and hardware operating status in real time, execute AI optimization process in normal mode, and immediately switch to hard-coded emergency pass-through mode in extreme emergencies or hardware anomalies, forming a complete failure safety closed loop.

[0065] 2. The core method flow of this invention:

[0066] A method for real-time autonomous avoidance of space debris in orbit includes the following steps:

[0067] Step S1: Multi-source threat perception and status acquisition. The spacecraft receives collision warning information from at least one information source (such as ground control stations, space situational awareness networks, and onboard autonomous sensors), and acquires its own status information such as orbit, attitude, velocity, and remaining fuel.

[0068] S11, Normal Mode: When the estimated collision probability is higher than the first threshold (e.g., 1×10⁻⁶), -4 If the value is below the second threshold and the onboard FPGA / ASIC is operating normally, then the conventional optimization and avoidance process from step S2 to step S6 is executed.

[0069] S12, Emergency Mode: When the estimated collision probability is much higher than the second threshold (e.g., 1×10⁻⁶), the emergency mode is activated when the estimated collision probability is much higher than the first threshold. -3When an abnormality is detected in the operation of the onboard FPGA / ASIC (such as watchdog timer timeout or logic self-test failure), the system skips the optimization calculation of the lightweight AI model and directly calls and executes a pre-stored emergency evasion maneuver instruction sequence that has been fully verified on the ground from the radiation-resistant ROM storage unit integrated in the onboard FPGA / ASIC chip, i.e., the onboard memory.

[0070] Step S2: Construction and deployment of lightweight AI models with hardware solidification.

[0071] S21, in the ground phase, compresses the complex teacher model into a lightweight student model with less than 1M parameters and less than 4MB in size through knowledge distillation, structured pruning and 8-bit quantization techniques.

[0072] S22, the lightweight student model is burned into the logic unit of the onboard FPGA or ASIC chip, making it a physical circuit that can be hardware accelerated for inference, with a single parallel inference power consumption of less than 100mW (milliwatt level) and a time consumption of less than 50 milliseconds.

[0073] Step S3: Hybrid intelligent decision-making – generating evasion maneuver candidate trajectories.

[0074] S31, Call the basic avoidance strategy from the standard avoidance strategy library pre-stored on the satellite. The standard avoidance strategy library contains at least three basic templates: orbital lag maneuver, orbital advance maneuver, and orbital out-of-plane maneuver, which are constructed based on prior knowledge of orbital mechanics.

[0075] S32, the warning information and the self-state information are input into the lightweight AI model embedded in the FPGA / ASIC chip. The model performs on-orbit real-time fine-tuning of the maneuver parameters (such as pulse size, application timing, and direction) of the above basic strategy template to generate 3-5 candidate trajectories for evasion maneuvers with physical feasibility.

[0076] Step S4: Rapid Safety Verification. Based on a simplified orbital dynamics model (primarily considering J2 perturbation, atmospheric drag, and other major perturbation factors), a millisecond-level rapid forward simulation and safety verification are performed on the candidate trajectory. The constraints of the safety verification include at least the following:

[0077] Minimum distance from debris after avoidance > safety threshold (e.g., 5 km);

[0078] The total velocity increment required for the candidate trajectory is less than the spacecraft's current remaining fuel equivalent velocity increment.

[0079] The estimated attitude maneuver angle and angular velocity during the maneuver are within the stable control range of the onboard attitude control system.

[0080] Step S5: Optimal Trajectory Selection. From all candidate trajectories that have passed the safety check, select the trajectory with the optimal fuel consumption.

[0081] S6, the pulse maneuver control command is directly sent to the on-board propulsion system for execution to complete the autonomous evasive maneuver.

[0082] The process of the autonomous avoidance method is described in [link to flowchart]. Figure 1 .

[0083] 3. System hardware implementation:

[0084] An on-board space debris autonomous avoidance system for implementing the above method includes:

[0085] Onboard communication module: Responsible for receiving and parsing multi-source collision warning information. It integrates the satellite's conventional S-band / X-band telemetry and control receiver to receive space situation data (SSA) warning information from ground stations; it also reserves an inter-satellite link interface for future connection to receive warning information broadcast by other satellites.

[0086] Onboard Processing Module: As the system's brain, its core is a space-grade FPGA or ASIC chip (such as Xilinx Kintex UltraScale or a domestically produced equivalent). This chip provides sufficient logic units, DSP slices, and on-chip memory to run lightweight AI models and fast safety verification algorithms in parallel using hardware acceleration. The chip contains lightweight AI models with fewer than 1 million parameters and a file size of less than 4MB, and stores a standard avoidance strategy library and emergency avoidance command sequences. This module is electrically connected to the communication module and simultaneously obtains necessary spacecraft status information from other subsystems of the satellite platform (such as the Attitude and Orbit Control System (AOCS), power system, etc.) via the satellite's internal data bus (such as SpaceWire or CAN). This includes position, velocity, and attitude information from GNSS receivers and star sensors, as well as remaining fuel information from the satellite management system. This module is configured to independently complete the entire decision-making process from threat assessment to command generation without ground intervention and possesses dual-mode operation capabilities.

[0087] In the first mode, the lightweight AI model is run, and a conventional optimization and avoidance process is executed.

[0088] In the second mode, when the emergency triggering conditions are met, the lightweight AI model is bypassed, and the emergency evasion instruction sequence is directly invoked and executed.

[0089] This module also includes a model update interface, which is used to receive model update data packets transmitted via the ground telemetry and control link, and to reconfigure some parameters and upgrade the version of the lightweight AI model embedded in the FPGA / ASIC chip. During the satellite safety period, the model embedded in the FPGA is reconfigured online (partial reconfiguration) to realize the on-orbit upgrade of the algorithm.

[0090] Onboard Execution Module: Electrically connected to the processing module, including the propulsion controller and thrusters (such as xenon ion thrusters or cold gas thrusters). The propulsion controller connects to the onboard processing module via an I553B or CAN bus to receive and precisely execute maneuver control commands.

[0091] 4. Design, training, and on-orbit deployment of lightweight AI models:

[0092] The lightweight AI model described in this invention is prepared and deployed through the following systematic process, and its lifecycle is detailed in [link to relevant documentation]. Figure 3 .

[0093] (a) Ground training and compression process:

[0094] Teacher Model Training: At a ground-based data center, using high-performance GPU servers, a complex teacher model was trained on over 100,000 simulated threat scenarios based on the Proximal Policy Optimization (PPO) deep reinforcement learning algorithm. The training scenarios covered different orbital altitudes, debris relative velocities, approach angles, and initial phases. The reward function was carefully designed to comprehensively consider fuel consumption (-ΔV), miss distance (+Miss Distance), and attitude stability.

[0095] Knowledge distillation and structured pruning: Using knowledge distillation (distillation temperature coefficient T=5), the decision-making ability of the teacher model is transferred to a minimally structured student model (e.g., a fully connected network with only 3 hidden layers and 128 nodes per layer). Subsequently, the student model undergoes up to 60% structured pruning, removing neural connections that contribute little to the output.

[0096] 8-bit integer quantization: The pruned model weights and activation values ​​are quantized into 8-bit integers (INT8Quantization), which converts the original 32-bit floating-point parameters into 8-bit integers, greatly reducing the model size and computational overhead.

[0097] (b) Lightweight effect and target specifications:

[0098] Following the above process, the resulting lightweight AI model has a parameter count and size that are reduced by one to two orders of magnitude compared to the original teacher model. For example, a teacher model with an initial parameter count of approximately 12M and a size of approximately 52MB can be optimized using this method to achieve a target size of <1M parameters and <4MB model size. The quantized model's trajectory planning performance loss on the test set is strictly controlled within 3%, fully meeting the accuracy requirements for on-orbit avoidance.

[0099] (c) On-orbit deployment:

[0100] The resulting lightweight AI model file (.bin or .bit format) is compiled using a dedicated toolchain and directly burned into the logic cells and on-chip memory of the FPGA. The model runs on the FPGA in hardware parallel acceleration in the form of a feedforward inference network, with a single parallel inference power consumption of less than 100mW (milliwatt level) and a time consumption of less than 50 milliseconds.

[0101] 5. Construction of a standard evasion strategy library:

[0102] The standard evasion strategy library pre-stored on the satellite is stored in a parameterized form and includes at least:

[0103] Track hysteresis maneuver: characterized by applying a negative pulse velocity increment (-ΔV_t) in the track tangential direction;

[0104] Orbital advance maneuver: Characterized by applying a positive pulse velocity increment (+ΔV_t) in the orbital tangential direction;

[0105] Out-of-plane maneuver: characterized by applying a pulse velocity increment (ΔV_n) in the orbital normal.

[0106] The task of lightweight AI models is to rapidly fine-tune and combine the parameters of these basic strategies (such as the magnitude of ΔV and the timing of its application) in orbit to generate candidate trajectories that adapt to specific threat scenarios.

[0107] The technical solution of the present invention will be described in detail below with reference to a preferred embodiment using a low-Earth orbit remote sensing satellite as a platform. It should be noted that this embodiment is only used to illustrate the core idea of ​​the present invention and is not the only limitation thereof.

[0108] Example: Low Earth Orbit Satellite Autonomous Avoidance System Based on Software and Hardware Collaboration.

[0109] (1) System hardware configuration:

[0110] The autonomous avoidance system in this embodiment is an intelligent functional unit of the satellite platform. Its hardware composition mainly includes the following three core modules. For the system composition and data flow, please refer to [link to documentation]. Figure 2 :

[0111] (a) Onboard communication module: integrates the satellite's conventional S-band / X-band telemetry and control receiver to receive space situation data (SSA) early warning information from ground stations; reserves an interface for future access to inter-satellite links to receive early warning information broadcast from other satellites.

[0112] (b) Onboard Processing Module: The core chip uses an aerospace-grade FPGA (such as a Xilinx Kintex UltraScale or a domestically produced equivalent). This chip provides sufficient logic units, DSP slices, and on-chip memory to run lightweight AI models and fast safety verification algorithms in parallel using hardware acceleration. The chip contains a lightweight AI model with fewer than 1 million parameters and a file size of less than 4 MB, and stores a standard avoidance strategy library and emergency avoidance command sequences. This module obtains necessary spacecraft status information from other subsystems of the satellite platform via the satellite's internal data bus (such as SpaceWire or CAN), including position, velocity, and attitude information from GNSS receivers and star sensors, as well as remaining fuel information from the satellite management system.

[0113] (c) Onboard Execution Module: The satellite's original propulsion controller and thrusters (such as xenon ion thrusters or cold gas thrusters). The propulsion controller is connected to the onboard processing module via a 1553B or CAN bus to receive and precisely execute maneuver control commands.

[0114] (2) Autonomous avoidance process:

[0115] Once the system receives a collision warning (including threat fragment ID, orbital elements, collision probability, nearest approach point TCA, etc.) via the communication module, it initiates the following autonomous closed-loop process:

[0116] Step S1: The onboard processing module receives early warning information through the onboard communication module, and simultaneously obtains the satellite's current position, velocity, attitude, and remaining fuel status from other subsystems on the platform via the satellite's internal data bus. The received and acquired information is then input into the onboard hardware acceleration computing platform for real-time collision probability estimation.

[0117] When the estimated collision probability is higher than the first threshold, the conventional optimization avoidance process of steps S2 to S6 is executed.

[0118] When the estimated collision probability is higher than the second threshold, the second threshold is greater than the first threshold, or an abnormal operation of the onboard hardware acceleration computing platform is detected, an emergency response procedure is executed.

[0119] Step S2: Construct a unified input feature vector by fusing the early warning information and the user's own state information, and then feed it into the lightweight AI model embedded in the FPGA.

[0120] Step S3: The lightweight AI model outputs 3-5 diverse evasion maneuver candidate trajectories in parallel within 50 milliseconds (each trajectory includes the ΔV vector and the maneuver time).

[0121] Step S4: All candidate trajectories are fed into a fast verifier based on a simplified orbital dynamics model. This simplified model retains key influencing factors such as Earth J2 perturbation and atmospheric drag perturbation (applicable to low Earth orbit), and completes parallel forward simulation and the following verification for all trajectories within 100 milliseconds while ensuring physical accuracy:

[0122] Minimum miss distance > 5 km (configurable);

[0123] Total speed increment Σ|ΔV| < remaining fuel equivalent ΔV;

[0124] The estimated attitude maneuver angles and angular velocities during the maneuver process are within the stable range of the control system.

[0125] Only tracks that pass all verifications can proceed to the next stage.

[0126] In this embodiment, the minimum safe distance threshold for the miss distance is set to 5 kilometers, based on the guidelines of the International Space Debris Coordination Committee (IADC) and engineering experience regarding typical low-Earth orbit spacecraft trajectory determination errors. This value is much larger than the uncertainty of the spacecraft's dimensions and orbit prediction, ensuring that the probability of collision is reduced to an acceptablely low level (typically <1×10⁻⁶). -7 This setting also strikes a good balance between safety and precious onboard fuel consumption. This threshold is a configurable parameter and can be adjusted according to the specific mission's safety requirements and orbital characteristics.

[0127] Step S5: The system selects the trajectory with the minimum fuel consumption (Σ|ΔV|) from the calibrated trajectories as the final execution strategy and generates the corresponding pulse control command.

[0128] S6 sends pulse maneuver control commands directly to the onboard propulsion system for execution, enabling autonomous evasive maneuvers.

[0129] Step S1 includes two execution modes:

[0130] ① Normal mode: (Steps S2 to S6) Generate pulse control commands based on the optimal trajectory, send them to the on-board propulsion system for execution, and complete autonomous evasion maneuvers.

[0131] ② Emergency Mode: When the estimated collision probability is higher than the second threshold or the FPGA watchdog timer times out or the logic self-test fails, the system immediately bypasses the lightweight AI model and directly calls and executes a preset, conservative out-of-plane emergency avoidance instruction from the policy library (for example, applying a fixed normal ΔV) to ensure that the maneuver is initiated within seconds, trading decision speed for survival probability.

[0132] The finalized maneuver control commands are sent directly to the propulsion controller of the onboard execution module. The controller then generates a specific thruster ignition command sequence and executes it immediately to complete the evasive maneuver. After the maneuver, a safety status self-assessment is initiated. The actual trajectory after the evasion is measured using the onboard GNSS receiver or inter-satellite link and compared with the expected safe trajectory to confirm the avoidance of collision risk.

[0133] (3) Specific implementation of emergency response mode:

[0134] In this embodiment, the first threshold (normal mode) is set to a collision probability Pc = 1 × 10⁻⁶. -4 (This probability level typically corresponds to the "red" alert in international standards for initiating evasive decision-making.) When Pc exceeds this value, the complete S2-S6 process described above is activated. The second threshold (emergency mode) is set to a collision probability Pc = 1 × 10⁻⁶. -3 (This indicates an extreme emergency situation with extremely limited onboard response time.) The system monitors collision probability and FPGA operating status in real time (through periodic logic self-tests and watchdog timers). When a second threshold trigger or hardware anomaly is detected, the system completes mode switching within milliseconds (<10 milliseconds) and directly executes hard-coded instructions.

[0135] (4) Model update mechanism:

[0136] After ground-based R&D personnel develop a new, lightweight model with superior performance, they package it into an encrypted data packet and upload it to the satellite via the telemetry, tracking, and command (TT&C) link. The onboard processing module's model update interface receives and verifies the data packet, and within the satellite's safe operating hours, performs online partial reconfiguration on the model embedded in the FPGA, enabling on-orbit algorithm upgrades.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time autonomous avoidance of space debris in orbit, characterized in that, Includes the following steps: S1 receives space debris collision warning information, acquires the spacecraft's own status information, and inputs it into the onboard hardware acceleration computing platform to estimate the collision probability in real time. When the estimated collision probability is higher than the first threshold, the conventional optimization avoidance process of steps S2 to S6 is executed. When the estimated collision probability is higher than the second threshold, the second threshold is greater than the first threshold, or an abnormal operation of the onboard hardware acceleration computing platform is detected, an emergency response procedure is executed: skip the optimization calculation of the lightweight AI model deployed on the onboard hardware acceleration computing platform, and directly call and execute the pre-stored emergency evasion maneuver instruction sequence from the onboard memory. S2, input the warning information and the self-state information into the lightweight AI model; The self-state information includes the spacecraft's orbit, attitude, velocity, and remaining fuel status information; The lightweight AI model is embedded in the logic unit of the onboard hardware-accelerated computing platform. Specifically, the lightweight AI model is a miniature neural network obtained by compressing a teacher model, which has undergone a ground-based compression process including knowledge distillation, structured pruning, and 8-bit quantization. Its parameter count is less than 1M and its model size is less than 4MB. The onboard hardware-accelerated computing platform is an FPGA chip or an ASIC chip, and the lightweight AI model is embedded in the logic unit of the chip. S3, using the lightweight AI model and combining it with the standard evasion strategy library pre-stored in the onboard memory, at least one evasion maneuver candidate trajectory is generated in real time on orbit. The standard avoidance strategy library includes at least three basic strategy templates based on prior knowledge of orbital mechanics: orbital lag maneuver, orbital advance maneuver, and out-of-plane maneuver. The standard avoidance strategy library is pre-stored in parameterized form in the radiation-resistant ROM storage unit integrated in the onboard FPGA chip or ASIC chip. The step of generating candidate trajectories further includes: fine-tuning the maneuver parameters of the basic strategy template in real time on orbit using the lightweight AI model; S4. Based on a simplified orbital dynamics model, perform rapid forward simulation and safety verification on the at least one candidate trajectory for evasion maneuver. The constraints of the safety verification include at least the minimum miss distance from the debris after evasion, total fuel consumption, and attitude stability margin during the maneuver. S5: Select the optimal trajectory from the candidate trajectories after safety verification and generate the corresponding pulse maneuver control command; S6, the pulse maneuver control command is directly sent to the on-board propulsion system for execution to complete the autonomous evasive maneuver.

2. The method according to claim 1, characterized in that, The specific constraints of the security verification in step S4 are as follows: The minimum distance from the debris after avoidance is greater than a preset safety threshold; the total velocity increment required for the avoidance maneuver candidate trajectory is less than the spacecraft's current remaining fuel equivalent velocity increment; and the estimated attitude maneuver angle and angular velocity during the maneuver are within the stable control range of the onboard attitude control system.

3. A space debris real-time autonomous avoidance system for implementing the method of any one of claims 1 to 2, characterized in that, include: The onboard communication module is used to receive early warning information about multi-source collisions with space debris; The onboard processing module is electrically connected to the onboard communication module, and its core processing unit is an FPGA chip or an ASIC chip. The FPGA chip or ASIC chip has a lightweight AI model with fewer than 1M parameters and a size of less than 4MB burned into it, and stores a standard avoidance strategy library and an emergency avoidance instruction sequence. The onboard processing module is configured to independently complete the entire decision-making process from threat assessment to command generation without ground intervention, and has dual-mode operation capability: In the first mode, the lightweight AI model is run, and the conventional optimization and avoidance process of steps S2 to S6 is executed. In the second mode, when the triggering conditions of the emergency response process in step S1 are met, the lightweight AI model is bypassed, and the emergency evasion maneuver instruction sequence is directly invoked and executed. The onboard execution module, electrically connected to the onboard processing module, includes a propulsion controller and a thruster, and is used to receive and execute the maneuver control commands.

4. The system according to claim 3, characterized in that: The onboard processing module also includes a model update interface, which is used to receive model update data packets transmitted via the ground telemetry and control link, and to reconfigure some parameters and upgrade the version of the lightweight AI model embedded on the FPGA or ASIC chip.

5. The system according to claim 3, characterized in that: After completing the evasive maneuver, the system initiates a self-assessment of its safety status, using an onboard GNSS receiver or inter-satellite link to measure the actual trajectory after the evasion and compare it with the expected safe trajectory to confirm the risk of avoiding collision.

6. A spacecraft, characterized in that: It is equipped with a space debris real-time autonomous avoidance system as described in any one of claims 3 to 5.

Images