Space multi-target on-orbit behavior prediction method and system based on large model reasoning

Abstract

The invention discloses a space multi-target on-orbit behavior prediction method and system based on large model reasoning, and the method comprises the steps: building a multi-level simulation scene based on an orbit dynamics model, simulating a plurality of space motions, and generating multi-dimensional time series data; single-target behavior data and multi-target interaction behavior data are obtained, high-level semantic feature mapping is carried out based on defined multi-class space target core intentions and behavior-intention association rules, a priori corpus knowledge base is obtained, and an attribute graph model is adopted for storage and updating; under the dual drive of'data-knowledge ', time sequence state change and topological relation evolution are processed at the same time through a fine-tuned large model, orbit dynamic constraints are embedded, and future behavior probability distribution and intention classification confidence are output. According to the method, a physical rule is creatively embedded into an intelligent reasoning process, probability distribution prediction and confidence evaluation of multi-target collaboration / confrontation behaviors are supported, and space situation awareness and on-orbit task planning are effectively supported.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology for spacecraft, and in particular to a method and system for predicting the on-orbit behavior of multiple space targets based on large model reasoning. Background Technology

[0002] The rapid development of satellite constellations and cluster architectures is essentially an inevitable choice to address the ever-increasing complexity of on-orbit missions. With the integration of diverse on-orbit needs such as remote sensing, global communication, and emergency response, the traditional isolated, single-mission operation mode can no longer meet the requirements of a highly dynamic and real-time space service system, driving the evolution of aerospace systems towards distributed collaborative clusters. However, while the expansion of cluster size and the tight coupling of functions enhance service capabilities, they also expose two core characteristics: the unknown behavior of on-orbit targets and the coupling of multi-target collaborative intentions. The former manifests as the unpredictability of multi-agent behavior in a dynamic space environment—when spacecraft respond to orbital perturbations, sudden commands, or resource competition, their autonomous decisions may result in unpredictable behaviors due to local environmental disturbances or global information lags. Such individual behavioral deviations, propagated through inter-satellite interaction networks, can easily trigger a chain reaction of instability in the overall mission logic of the cluster. The latter stems from the deeply intertwined nature of multi-objective collaboration. When a cluster needs to simultaneously support differentiated tasks such as navigation and positioning enhancement, space environment monitoring, and on-orbit service assurance, the resource requirements and execution logic constraints of each objective may create implicit conflicts. This coupling effect at the intent level exacerbates the uncertainty of the system state, leading to a decline in global performance. Resolving such contradictions requires constructing a collaborative paradigm at the architectural level that features dynamic objective perception, intent decoupling mapping, and elastic resource isolation. This can be achieved through a multi-level closed-loop feedback mechanism to enable autonomous adaptation and robust evolution of complex task flows.

[0003] Current space target behavior prediction technologies primarily rely on physical model-driven methods, which suffer from drawbacks such as limited modeling dimensions and lagging real-time response in tasks like target intent recognition and abnormal behavior detection. With the breakthrough capabilities demonstrated by advanced large-scale language models in cross-modal representation learning and temporal relationship modeling, their powerful inference performance provides a new paradigm for space behavior sequence analysis. Existing large-scale models are mainly trained on general scenarios and large-scale data. However, the unique operating environment of spacecraft, including dynamic orbital changes, complex space mechanics, and long-distance communication with Earth, results in data patterns that differ significantly from the conventional data patterns of large-scale models. Existing models struggle to accurately capture and understand the unique patterns and on-orbit behavior of spacecraft. Furthermore, data generated during spacecraft operation often contains noise and uncertainty, making it difficult for large-scale models to process. Therefore, how to adaptively modify the framework of existing models to effectively handle the unique data of the spacecraft domain and accurately predict behavior is a pressing technical challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a spatial multi-target on-orbit behavior prediction method and system based on large model reasoning, construct a "data-knowledge" dual-model training library for on-orbit target behavior prediction and intent reasoning tasks, and achieve effective prediction and reasoning of complex spatial multi-target behavior intents by learning spatial on-orbit constraints.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for predicting the on-orbit behavior of multiple targets in space based on large model inference, comprising the following steps: S1, On-orbit Scenario Simulation: Based on the orbital dynamics model, a multi-level simulation scenario is constructed to simulate various space motions and generate multi-dimensional time-series data including basic state quantities, anomaly monitoring quantities, and event marker quantities; among them, the multi-level simulation scenario includes simulation scenarios of the main star sensing single and multiple targets; S2, Prior Corpus Knowledge Base Construction: Acquire single-target behavior data and multi-target interaction behavior data, perform high-level semantic feature mapping based on the defined multi-class spatial target core intent and behavior-intent association rules, obtain the prior corpus knowledge base, and store and update it using an attribute graph model; S3, On-orbit Behavior Prediction Analysis: Based on the multi-dimensional time-series data in S1 and the dynamic interactive topology graph in the attribute graph model in S2, under the dual drive of "data-knowledge", the large model with fine-tuning simultaneously processes the temporal state changes and topological relationship evolution, embeds orbital dynamic constraints, and outputs the probability distribution of future behavior and the confidence level of intention classification.

[0006] Furthermore, in S1, the method for constructing and initializing the orbital dynamics model is as follows: Based on the target characteristics and environmental parameters, an orbital dynamics model is constructed. The numerical integration method is used to solve a set of differential equations that include the Earth's non-spherical gravity, atmospheric drag, solar radiation pressure and third-body gravity. The target orbital parameters are initialized, the coordinate system is set and the time reference is calibrated to provide a unified multi-target spatiotemporal reference.

[0007] Furthermore, in S1, the multi-level simulation scenario specifically includes: One-to-one perception scenario: Establish a relative motion model between the main star and a single target, configure the field of view, resolution and sampling frequency of the lidar and star sensor, and simulate the relative orbit evolution in fixed-point observation or tracking mode; One-to-two and one-to-three perception scenarios: For debris swarms or satellite formations, establish a group orbit coupling model, use a probability density function to describe the target distribution, and calculate the probability of approach events in real time through a collision cone algorithm.

[0008] Furthermore, in S1, various space motions include satellite formation coordination, debris drift, and non-cooperative target maneuvering; The multidimensional time-series data includes basic state quantities such as position and velocity vectors in the J2000 coordinate system and orbital six-axis numbers; abnormal monitoring quantities such as attitude angle deviation exceeding limits, propellant leakage rate, and engineering parameters of abnormal temperature of onboard computer; and event marker quantities such as maneuver identification markers, collision risk levels, and sensor failure codes.

[0009] Furthermore, in step S2, acquiring single-target behavior data and multi-target interaction behavior data specifically includes: For a single target in orbit, label its basic behavior patterns, including orbit maintenance, Hohmann transfer, and attitude maneuver. Establish a behavior pattern classification tree and classify the behavior in orbit into three levels, including primary maneuver, secondary orbital maneuver, and tertiary emergency collision avoidance. For multi-target motion in space, collect interaction event data, including formation relative position maintenance, inter-satellite link establishment, adversarial game, and record the IDs of the interacting parties, interaction type, and threat distance.

[0010] Furthermore, in S2, based on the defined multi-class spatial target core intent and behavior-intent association rules, high-level semantic feature mapping is performed, specifically including: Define the core intents of multiple types of space targets, including reconnaissance, relay, escort, approach, departure, and evasion, and expand subclasses for each type of intent; Establish behavior-intent association rules, and label conflict type and conflict intensity in interaction event data to construct a conflict-intent association matrix and form a high-level semantic feature mapping.

[0011] Furthermore, in step S2, the storage and updating are performed using an attribute graph model, specifically including: An attribute graph model is used to store knowledge in the prior corpus knowledge base. The nodes in the attribute graph model include satellite entities, intent nodes, and conflict nodes. The edge relationships are defined in three categories: behavior-intent edge, entity-conflict edge, and intent-conflict edge. Based on a descriptive logic-based reasoning framework, it detects logical contradictions in the intent-behavior-conflict triplet; it automatically triggers subgraph updates when a new event occurs; and it establishes a semantic drift detection module to reconstruct the prior corpus knowledge base when the intent classification accuracy drops below a threshold.

[0012] Furthermore, in S3, the fine-tuned large model adopts a Transformer-GNN hybrid architecture; By fine-tuning a large model to simultaneously handle temporal state changes and topological evolution, embedding orbital dynamics constraints, and outputting a probability distribution of future behavior and an intention classification confidence level (not specifically including: The Transformer layer introduces orbital periodic phase encoding and energy conservation regularization to capture long-range temporal dependencies and constrain physical rationality; the GNN layer extracts relative motion features between targets through relational perception aggregation and integrates spatiotemporal features with cross-attention mechanism; the output adopts a hybrid density network to generate maneuver probability distribution and intention classification confidence, and quantifies cognitive uncertainty based on Dirichlet distribution.

[0013] Furthermore, the method also includes S4, dynamic evaluation of multi-target cooperative behavior prediction results: verifying physical consistency using orbital integral inversion, the system monitors the cooperative performance indicators of the formation satellites in real time, including: relative position deviation, prediction delay, and intention accuracy, and constructs a global intention-local behavior dual-channel evaluation framework to form a closed-loop perception link of "prediction-evaluation".

[0014] This invention also provides a space multi-object on-orbit behavior prediction system based on large model inference, applied to the above-mentioned space multi-object on-orbit behavior prediction method based on large model inference, comprising: The on-orbit scenario simulation module is used to construct multi-level simulation scenarios based on the orbital dynamics model, simulate various space motions, and generate multi-dimensional time-series data including basic state quantities, anomaly monitoring quantities, and event marker quantities; among them, the multi-level simulation scenarios include simulation scenarios of the main star sensing single targets and multiple targets. The prior corpus knowledge base construction module is used to acquire single-target behavior data and multi-target interaction behavior data. Based on the defined core intents of multiple spatial targets and behavior-intent association rules, high-level semantic feature mapping is performed to obtain the prior corpus knowledge base, which is stored and updated using an attribute graph model. The on-orbit behavior prediction and analysis module is used to process the temporal state changes and topological relationship evolution simultaneously through a fine-tuned large model driven by "data-knowledge" and based on multi-dimensional time series data and dynamic interactive topology graphs in attribute graph models. It embeds orbital dynamic constraints and outputs the probability distribution of future behavior and the confidence level of intention classification.

[0015] According to specific embodiments provided by the present invention, the spatial multi-object on-orbit behavior prediction method and system based on large model inference provided by the present invention discloses the following technical effects: (1) A multi-level spatial intent semantic parsing framework was innovatively constructed. Through the explicit association mechanism between physical behavior features and the core intent of the target, cross-level reasoning from the underlying orbital dynamics behavior to complex interactive intent was realized. Spatial intent was classified in multiple granularities. The intent mapping rule base was dynamically constructed by combining orbital maneuver features, environmental disturbance parameters and interactive events, which effectively solved the problem of ambiguity in the recognition of composite intents in traditional methods. The coupling and dynamic evolution characteristics of intents in adversarial scenarios were modeled by the conflict-intent association matrix, which significantly improved the intent interpretation capability in complex spatial and complex task scenarios.

[0016] (2) A two-way driving mechanism of spatial behavior data and dynamic knowledge base is proposed to establish a closed-loop cognitive enhancement system from multi-source heterogeneous data to domain knowledge; the deep integration of orbital state data, abnormal events and prior knowledge is realized through attribute graph model, and the logical consistency of behavior-intention is verified in real time by using logical reasoning engine, breaking through the limitations of traditional single-modal data analysis; the architecture supports the dynamic evolution and semantic self-correction of knowledge graph, can adapt to changes in spatial situation, and significantly enhances the cognitive robustness and interpretability of sudden spatial events.

[0017] (3) Deeply explore the core advantages of large models in complex spatiotemporal modeling, and propose a pre-training-fine-tuning framework based on prior knowledge of orbital dynamics; use self-attention mechanism to capture the cross-scale correlation between long-period orbital evolution law and short-term maneuvering behavior; design a prompting fine-tuning strategy guided by orbital physics, encode prior knowledge such as the non-spherical gravitational perturbation of the Earth and the gravitational influence of the third body into structured prompting vectors, and constrain the model to follow basic physical laws when generating prediction results; combine dynamic graph neural network to construct a multi-objective collaborative-adversarial relationship reasoning module, and analyze the implicit logic such as debris diffusion threat propagation path and formation satellite game strategy through hierarchical attention weight visualization technology; this method breaks through the modeling bottleneck of traditional models for complex space interaction scenarios, significantly improves the prediction ability of nonlinear dynamics such as concealed maneuvering intentions and chain collision risks, and supports zero-sample adaptive reasoning across mission scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the spatial multi-object on-orbit behavior prediction method based on large model inference of the present invention; Figure 2 This is a schematic diagram of the spatial multi-objective on-orbit behavior prediction framework based on large model inference of the present invention; Figure 3 This is a schematic diagram illustrating the multi-target hierarchical intent division of the present invention; Figure 4 This is a flowchart of the on-orbit constraint process for fine-tuning the large model of this invention. Figure 5 This is a schematic diagram of the target behavior prediction results of the present invention. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. These embodiments are intended to explain the present invention and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Materials or instruments whose manufacturers are not specified are all conventional products that can be obtained commercially.

[0021] This invention innovatively embeds physical laws into the intelligent reasoning process, supporting probability distribution prediction and confidence assessment of multi-target cooperative / adversarial behaviors, effectively supporting space situational awareness and on-orbit mission planning.

[0022] The present invention provides a method for predicting the on-orbit behavior of multiple targets in space based on large model inference, comprising the following steps: I. On-orbit Scenario Simulation (Orbital Dynamics, Motion Characteristics Simulation): Combining orbital dynamics models, we construct simulation scenarios with one-to-one, one-to-two, and one-to-three perception capabilities to simulate typical motions such as satellite formation coordination, debris drift, and non-cooperative target maneuvers. This generates multi-dimensional time-series data containing time, position, abnormal states, and event markers, providing data support for predicting the behavior of multiple targets in space, solving the problem of result bias caused by data sparsity, and supporting the probability calculation of scenario threat situations and the deduction of maneuver intentions.

[0023] II. Construction of Prior Corpus Knowledge Base (Definition of Individual Behavioral Intent and Multi-Agent Coupled Behavioral Intent): A corpus knowledge base covering individual behaviors and multi-agent interactions (such as cooperative relay and adversarial games) is constructed. Physical motion data is associated with high-level semantic features, and communication, threat, and dependency relationships between targets are represented. Six types of spatial intent are defined. Through a dual-driven "data + knowledge" model, the large model's ability to autonomously understand complex spatial tasks is supported, enabling it not only to predict target trajectories but also to analyze the tactical intent behind the behaviors.

[0024] III. On-orbit Behavior Prediction and Analysis (Large Model Fine-tuning Training, Space Multi-Objective Behavior Prediction, Intent Inference): Multi-objective behavior prediction based on fine-tuned large models achieves a deep integration of physical laws and intelligent reasoning. By embedding orbital dynamics constraints into a Transformer-GNN hybrid architecture, the model can simultaneously handle temporal state changes and topological evolution, outputting the probability distribution and confidence level of future behaviors. A two-stage training strategy (pre-training + online learning) balances the differences between training data and online operation, overcoming the limitations of traditional single-objective independent prediction, dynamically evaluating the chain reaction of multi-objective cooperation / adversarial actions, and providing real-time, interpretable complex intent inference capabilities for tasks such as space attack and defense decision-making and on-orbit service planning.

[0025] Specifically, such as Figure 1 and Figure 2As shown in the embodiments of the present invention, the space multi-target on-orbit behavior prediction method based on large model inference is mainly aimed at space on-orbit situational awareness tasks. The overall process includes multiple stages such as spacecraft on-orbit scenario simulation, construction of a space prior knowledge base, on-orbit behavior prediction analysis based on large model inference, and dynamic evaluation of multi-target collaborative behavior prediction results, forming a closed loop of space target on-orbit perception and understanding driven by both "data and knowledge". Specifically, it includes the following steps: Step (1), Orbital Dynamics Modeling Initialization: Based on target characteristics such as mass and inertia, and environmental parameters, a high-precision orbital dynamics model is constructed. The Runge-Kutta fourth-order algorithm is used to solve the differential equations involving Earth's non-spherical gravity, atmospheric drag, solar radiation pressure, and third-body gravity. The target orbital parameters are initialized, the coordinate system is set, and the time reference is calibrated to ensure the consistency of the spatiotemporal reference for multiple targets.

[0026] Step (2), Multi-level simulation scenario construction: ① One-to-one perception scenario: Establish a relative motion model between the main star and a single target, configure the field of view, resolution and sampling frequency of the lidar / star sensor, and simulate the relative orbit evolution in fixed-point observation or tracking mode; ② One-to-two and one-to-three perception scenarios: For debris groups or satellite formations, establish a group orbit coupling model, use a probability density function to describe the target distribution, and calculate the probability of approach events in real time through the collision cone algorithm; Step (3), Multidimensional Time Series Data Generation and Structure: Generate a data stream in the virtual scene with microsecond-level time synchronization accuracy, including: basic state quantities: position / velocity vectors in the J2000 coordinate system, orbital six roots; anomaly monitoring quantities: attitude angle deviation exceeding limits, propellant leakage rate, abnormal temperature of onboard computer and other engineering parameters; event flag quantities: maneuver identification flag (0 / 1), collision risk level (level 1-5), sensor failure code.

[0027] Step (4), multi-granularity behavior data collection and structuring: For single targets in orbit, label basic behavior patterns such as orbit maintenance, Hohmann transfer, attitude maneuver, etc., establish a behavior pattern classification tree, and classify the in-orbit behavior into three levels, such as Level 1 "maneuver" → Level 2 "orbit maneuver" → Level 3 "emergency collision avoidance"; For multi-target motion in orbit, collect data on interaction events such as formation relative position maintenance, inter-satellite link establishment, and adversarial game, and record the IDs of the interacting parties, interaction type, and threat distance.

[0028] Step (5), High-level semantic feature mapping: Define a spatial intent identification framework, including 6 core intents (reconnaissance, relay, escort, approach, departure, and avoidance), and expand each intent into subclasses, such as avoidance → collision avoidance / electromagnetic silence avoidance; further establish physical behavior-semantic intent association rules: ① Rule engine, when a ΔV pulse is detected pointing outward from the orbital plane and lasting for more than 10 seconds, it is mapped to "orbital plane adjustment → avoidance intent"; ② Statistical association: through frequent itemset mining, a strong correlation is found between "inter-satellite link establishment frequency > 5 times / hour" and "cooperative relay intent". In the interaction event data, label the conflict type, conflict intensity (0-1 continuous value), and resolution strategy (negotiation / coercion / concession); construct a conflict-intent association matrix, for example, "orbital resource preemption conflict" may be associated with a combination of "stationing intent" and "interception intent".

[0029] Through steps (4) and (5), the following is formed: Figure 3 The multi-objective hierarchical intent classification shown is divided into single-objective behavioral intent, dual-objective behavioral intent, and multi-objective behavioral intent. Figure 3 The definition of "behavior-intention" is carried out from several perspectives, including orbital calculation for individual targets and inter-satellite relative state analysis for multiple targets. As the number of targets increases, the association knowledge supporting the evaluation of the collaborative intention of multiple targets is constructed based on the analysis of the motion intention of individual targets.

[0030] Step (6) Corpus Verification and Iteration: Attribute graph model is used to store knowledge. Nodes include satellite entities (ID, type, affiliated institution), intent nodes (type, confidence level), and conflict nodes (type, urgency level). Three types of edge relationships are defined: ① Behavior-Intent Edge: Weighted edge (weight = support of behavior for intent); ② Entity-Conflict Edge: Dynamic edge with timestamp, representing the conflict lifecycle; ③ Intent-Conflict Edge: Directed edge, indicating the causal direction of intent initiating or mitigating conflict. Develop an inference engine based on descriptive logic to detect logical contradictions in intent-behavior-conflict triples; automatically trigger subgraph updates when new space-on-orbit events occur; establish a semantic drift detection module to reconstruct the constructed prior knowledge corpus when the intent classification accuracy drops below a threshold.

[0031] Step (7) Large model fine-tuning training and learning: such as Figure 4 As shown, the target's orbital state time series data and dynamic interaction topology graph are input into the model respectively. Orbital period phase encoding and energy conservation regularization are introduced in the Transformer layer to capture long-range temporal dependencies and constrain physical rationality. The Graph Neural Network (GNN) layer extracts the relative motion features between targets through relational perception aggregation and combines spatiotemporal features with a cross-attention mechanism. The output uses a hybrid density network to generate the maneuver probability distribution and intention classification confidence, and quantifies cognitive uncertainty based on the Dirichlet distribution.

[0032] Specifically, in the composite model for collaborative perception and intent reasoning of formation satellites, a dual-channel input architecture is constructed to process heterogeneous data sources. Orbital state time-series data is processed using an improved Transformer encoder, where orbital period phase encoding is implemented through Fourier series expansion: given a timestamp... and orbital period Phase features are calculated and superimposed with standard position codes to form a spatiotemporally aware position embedding. To ensure the rationality of the dynamics, an energy conservation regularization term is introduced into the self-attention calculation:

[0033] in, It is a velocity vector. For position vectors, The gravitational constant, Nominal orbital energy, represents the weighting coefficients of the energy conservation constraint, which is incorporated into the loss function using the Lagrange multiplier method.

[0034] A dynamic interactive topology graph is input into a relation-enhanced graph neural network, where edge weights in the adjacency matrix are defined as functions of relative motion parameters. A multi-head relation-aware message passing mechanism is employed, with attention coefficients dynamically adjusted by the derivative of relative distance. In the spatiotemporal feature fusion stage, a cross-modal cross-attention mechanism is designed to achieve collaborative interaction of spatiotemporal features. GNN node features are used as queries, and Transformer temporal features are used as key-value pairs to calculate cross-domain attention weights.

[0035] in, The query vector comes from the GNN node. K k The key vector is derived from the temporal features of the Transformer. d The dimension of the Query / Key vector, used as a scaling and normalization factor; like Figure 5As shown, in a specific embodiment, the orbital state of the target spacecraft was dynamically analyzed: First, regarding orbital dynamics, the velocity in the y-direction showed a continuous decreasing trend (from 0.8968 to 0.8836), while the velocity in the x-direction continuously increased (from 0.4214 to 0.4427), reflecting that the spacecraft was experiencing a coupled effect of tangential acceleration and radial deceleration. Simultaneously, the velocity in the z-direction continuously increased (from 0.1346 to 0.1526), ​​but the displacement gradually approached zero, suggesting that the spacecraft was undergoing orbital plane adjustment, with its orbital inclination changing slowly. Second, in orbital type determination, the semi-major axis and eccentricity obtained through inversion of instantaneous position and velocity vectors indicated that the spacecraft was in a near-circular, low-eccentricity elliptical orbit. Furthermore, regarding perturbations and stability, the continuous increase in the z-direction velocity was inferred to be an orbital plane correction maneuver, which compensated for external disturbances through normal thrust to maintain orbital stability. Finally, at the level of the overall target intent, the probability of the target spacecraft approaching our spacecraft is as high as 94.56%. The solid line in the predicted trajectory diagram corresponds to this high-probability event, while the dashed line indicates another possible path with a suboptimal probability of 62.94%.

[0036] Step (8) Dynamic evaluation of multi-target cooperative behavior prediction results: Use orbital integral inversion to verify physical consistency. The system monitors the cooperative performance indicators of the formation satellites in real time: relative position deviation, prediction delay, intention accuracy, etc., and constructs a global intention-local behavior dual-channel evaluation framework to form a closed-loop perception link of "prediction-evaluation".

[0037] Specifically, the output employs a hybrid density network to generate the probability distribution of maneuvering behavior, while simultaneously using a Dirichlet distribution to model and classify uncertainties. At the physical verification level, an orbital integral inversion module is established to solve the perturbed motion equations. A fourth-order Runge-Kutta method is used for numerical integration, and the residual norm between the predicted state values ​​and the actual observations is calculated as a physical consistency index.

[0038] Multi-dimensional performance evaluation system: Relative position deviation is calculated using Mahalanobis distance.

[0039] Where M measures the statistical deviation between the predicted and the actual location. This is an estimate of the orbital position output by the model; Predicted latency metrics:

[0040] in, The latency is the time required for a single inference attempt, and N is the total number of test samples. For the first i Predicted timestamps for each sample For the first i The actual timestamp of each sample; Intent classification accuracy is calculated using the confusion matrix:

[0041] in, P c For category c The true number of cases, where C is the total number of categories.

[0042] By monitoring these indicators online, a vector is formed. , To indicate the overall performance status, a dynamic feedback mechanism is established. The system triggers online updates of model parameters, forming a closed-loop perception enhancement link from spatiotemporal feature extraction to physical verification and system optimization; among which, Represents the reference performance vector. This indicates the trigger threshold; when the system performance deviates from the reference deviation, an online update of the model is initiated.

[0043] This invention also provides a space multi-object on-orbit behavior prediction system based on large model inference, applied to the above-mentioned space multi-object on-orbit behavior prediction method based on large model inference, comprising: The on-orbit scenario simulation module is used to construct multi-level simulation scenarios based on the orbital dynamics model, simulate various space motions, and generate multi-dimensional time-series data including basic state quantities, anomaly monitoring quantities, and event marker quantities; among them, the multi-level simulation scenarios include simulation scenarios of the main star sensing single targets and multiple targets. The prior corpus knowledge base construction module is used to acquire single-target behavior data and multi-target interaction behavior data. Based on the defined core intents of multiple spatial targets and behavior-intent association rules, high-level semantic feature mapping is performed to obtain the prior corpus knowledge base, which is stored and updated using an attribute graph model. The on-orbit behavior prediction and analysis module is used to process the temporal state changes and topological relationship evolution simultaneously through a fine-tuned large model driven by "data-knowledge" and based on multi-dimensional time series data and dynamic interactive topology graphs in attribute graph models. It embeds orbital dynamic constraints and outputs the probability distribution of future behavior and the confidence level of intention classification.

[0044] In summary, this invention discloses a method and system for predicting the on-orbit behavior of multiple targets in space based on large-scale model reasoning. It achieves accurate deduction of the behavioral intentions of multiple targets under complex space situations through a dual-mode "data-knowledge" driven mechanism. The method adopts a three-layer architecture: First, it constructs a spacecraft on-orbit scenario simulation module, generating multi-dimensional time-series simulation data covering satellite formation coordination, debris drift, and non-cooperative target maneuvers based on an orbital dynamics model, overcoming the sparsity limitations of actual observation data. Second, it establishes a space prior knowledge base, defining six types of space intentions and multi-body coupling behavior rules through the association mapping between physical motion data and high-level semantics. Finally, it designs a multi-target behavior prediction and analysis module based on a fine-tuned large-scale model, integrating orbital dynamics constraints and achieving joint modeling of temporal states and topological relationships through a two-stage training strategy. This invention innovatively embeds physical laws into the intelligent reasoning process, supporting probability distribution prediction and confidence assessment of multi-target cooperative / adversarial behaviors, effectively supporting space situational awareness and on-orbit mission planning.

[0045] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-described method for predicting the on-orbit behavior of multiple spatial targets based on large model inference.

[0046] Matters not covered in this invention are common knowledge.

[0047] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting the on-orbit behavior of multiple spatial targets based on large model inference, characterized in that, Includes the following steps: S1, On-orbit Scenario Simulation: Based on the orbital dynamics model, a multi-level simulation scenario is constructed to simulate various space motions and generate multi-dimensional time-series data including basic state quantities, anomaly monitoring quantities, and event marker quantities; among them, the multi-level simulation scenario includes simulation scenarios of the main star sensing single and multiple targets; S2, Prior Corpus Knowledge Base Construction: Acquire single-target behavior data and multi-target interaction behavior data, perform high-level semantic feature mapping based on the defined multi-class spatial target core intent and behavior-intent association rules, obtain the prior corpus knowledge base, and store and update it using an attribute graph model; S3, On-orbit Behavior Prediction Analysis: Based on the multi-dimensional time-series data in S1 and the dynamic interactive topology graph in the attribute graph model in S2, under the dual drive of "data-knowledge", the large model with fine-tuning simultaneously processes the temporal state changes and topological relationship evolution, embeds orbital dynamic constraints, and outputs the probability distribution of future behavior and the confidence level of intention classification.

2. The space multi-object on-orbit behavior prediction method based on large model inference according to claim 1, characterized in that, In S1, the method for constructing and initializing the orbital dynamics model is as follows: Based on the target characteristics and environmental parameters, an orbital dynamics model is constructed. The numerical integration method is used to solve a set of differential equations that include the Earth's non-spherical gravity, atmospheric drag, solar radiation pressure and third-body gravity. The target orbital parameters are initialized, the coordinate system is set and the time reference is calibrated to provide a unified multi-target spatiotemporal reference.

3. The method for predicting the on-orbit behavior of multiple spatial targets based on large model inference according to claim 1, characterized in that, In S1, the multi-level simulation scenario specifically includes: One-to-one perception scenario: Establish a relative motion model between the main star and a single target, configure the field of view, resolution and sampling frequency of the lidar and star sensor, and simulate the relative orbit evolution in fixed-point observation or tracking mode; One-to-two and one-to-three perception scenarios: For debris swarms or satellite formations, establish a group orbit coupling model, use a probability density function to describe the target distribution, and calculate the probability of approach events in real time through a collision cone algorithm.

4. The method for predicting the on-orbit behavior of multiple targets in space based on large model inference according to claim 1, characterized in that, In S1, various space motions include satellite formation coordination, debris drift, and non-cooperative target maneuvering; The multidimensional time-series data includes basic state quantities such as position and velocity vectors in the J2000 coordinate system and orbital six-axis numbers; abnormal monitoring quantities such as attitude angle deviation exceeding limits, propellant leakage rate, and engineering parameters of abnormal temperature of onboard computer; and event marker quantities such as maneuver identification markers, collision risk levels, and sensor failure codes.

5. The method for predicting the on-orbit behavior of multiple space targets based on large model inference according to claim 1, characterized in that, In step S2, acquiring single-target behavior data and multi-target interaction behavior data specifically includes: For a single target in orbit, label its basic behavior patterns, including orbit maintenance, Hohmann transfer, and attitude maneuver. Establish a behavior pattern classification tree and classify the behavior in orbit into three levels, including primary maneuver, secondary orbital maneuver, and tertiary emergency collision avoidance. For multi-target motion in space, collect interaction event data, including formation relative position maintenance, inter-satellite link establishment, adversarial game, and record the IDs of the interacting parties, interaction type, and threat distance.

6. The method for predicting the on-orbit behavior of multiple targets in space based on large model inference according to claim 1, characterized in that, In step S2, high-level semantic feature mapping is performed based on the defined multi-class spatial target core intent and behavior-intent association rules, specifically including: Define the core intents of multiple types of space targets, including reconnaissance, relay, escort, approach, departure, and evasion, and expand subclasses for each type of intent; Establish behavior-intent association rules, and label conflict type and conflict intensity in interaction event data to construct a conflict-intent association matrix and form a high-level semantic feature mapping.

7. The method for predicting the on-orbit behavior of multiple targets in space based on large model inference according to claim 1, characterized in that, In step S2, an attribute graph model is used for storage and updating, specifically including: An attribute graph model is used to store knowledge in the prior corpus knowledge base. The nodes in the attribute graph model include satellite entities, intent nodes, and conflict nodes. The edge relationships are defined in three categories: behavior-intent edge, entity-conflict edge, and intent-conflict edge. Based on a descriptive logic-based reasoning framework, it detects logical contradictions in the intent-behavior-conflict triplet; it automatically triggers subgraph updates when a new event occurs; and it establishes a semantic drift detection module to reconstruct the prior corpus knowledge base when the intent classification accuracy drops below a threshold.

8. The method for predicting the on-orbit behavior of multiple targets in space based on large model inference according to claim 1, characterized in that, In S3, the fine-tuned large model adopts a Transformer-GNN hybrid architecture; By fine-tuning a large model to simultaneously handle temporal state changes and topological evolution, embedding orbital dynamics constraints, and outputting a probability distribution of future behavior and an intention classification confidence level (not specifically including: The Transformer layer introduces orbital periodic phase encoding and energy conservation regularization to capture long-range temporal dependencies and constrain physical rationality; the GNN layer extracts relative motion features between targets through relational perception aggregation and integrates spatiotemporal features with cross-attention mechanism; the output adopts a hybrid density network to generate maneuver probability distribution and intention classification confidence, and quantifies cognitive uncertainty based on Dirichlet distribution.

9. The method for predicting the on-orbit behavior of multiple space targets based on large model inference according to claim 1, characterized in that, The method also includes S4, dynamic evaluation of multi-target cooperative behavior prediction results: using orbital integral inversion to verify physical consistency, the system monitors the cooperative performance indicators of the formation satellites in real time, including: relative position deviation, prediction delay, and intention accuracy, and constructs a global intention-local behavior dual-channel evaluation framework to form a closed-loop perception link of "prediction-evaluation".

10. A space multi-object on-orbit behavior prediction system based on large model inference, applied to the space multi-object on-orbit behavior prediction method based on large model inference as described in any one of claims 1-9, characterized in that, include: The on-orbit scenario simulation module is used to construct multi-level simulation scenarios based on the orbital dynamics model, simulate various space motions, and generate multi-dimensional time-series data including basic state quantities, anomaly monitoring quantities, and event marker quantities; among them, the multi-level simulation scenarios include simulation scenarios of the main star sensing single targets and multiple targets. The prior corpus knowledge base construction module is used to acquire single-target behavior data and multi-target interaction behavior data. Based on the defined core intents of multiple spatial targets and behavior-intent association rules, high-level semantic feature mapping is performed to obtain the prior corpus knowledge base, which is stored and updated using an attribute graph model. The on-orbit behavior prediction and analysis module is used to process the temporal state changes and topological relationship evolution simultaneously through a fine-tuned large model driven by "data-knowledge" dual-drive, based on multi-dimensional time series data and dynamic interactive topology graphs in attribute graph models. It embeds orbital dynamic constraints and outputs the probability distribution of future behavior and the confidence level of intention classification.

Images