Spacecraft intention reasoning optimization method based on multi-hop knowledge chain

Abstract

The invention discloses a spacecraft intention reasoning optimization method based on a multi-hop knowledge chain, and the method comprises the steps: firstly constructing a hierarchical knowledge framework containing spacecraft information, subsystems, components, behaviors and possible intentions, and extracting a dynamic sub-graph which fuses time sequence dependence and is associated with inherent semantics through a sliding window; the orbit dynamics constraint is embedded to ensure logic self-consistency; a hierarchical reasoning framework with rule guidance and graph neural network cooperation is adopted, an upper layer rapidly screens a high-confidence-coefficient intention candidate set based on predefined behavior pattern rules and orbital constraints, and a lower layer aggregates multi-hop semantic information along a knowledge chain through a lightweight graph neural network to generate fine-grained intention probability distribution; and finally, matching a historical behavior template through a dynamic time warping algorithm, and combining sequential logic constraint verification to realize intention dual verification. According to the method, the accuracy, the real-time performance and the engineering realizability of spacecraft behavior intention reasoning in a complex scene can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and aerospace technology, and in particular to an optimization method for spacecraft intent reasoning based on multi-hop knowledge chains. Background Technology

[0002] With the development of aerospace technology, the on-orbit operation of spacecraft is becoming increasingly complex, making accurate reasoning about their behavioral intentions crucial for ensuring the safe and efficient execution of space missions. Existing technologies employ traditional machine learning methods to analyze spacecraft behavioral data, but these methods struggle to effectively handle temporal dependencies and complex logical relationships within the data. This is particularly true when reasoning about behavioral intention chains involving multi-hop relationships, where a lack of structured knowledge can easily lead to reasoning breakdowns. While knowledge graph-based reasoning techniques can utilize graph structures to model entity relationships, existing solutions largely rely on general-domain multi-hop reasoning methods, such as random walks and path sorting algorithms, lacking specific modeling tailored to the characteristics of the spacecraft domain.

[0003] Specifically, multi-hop reasoning in the aerospace field involves multi-level associations and strong domain characteristics such as orbital dynamics constraints and temporal logical dependencies. However, traditional knowledge graph multi-hop reasoning faces key bottlenecks when dealing with complex spacecraft behavior chains: on the one hand, it fails to deeply integrate the physical laws of the aerospace field with the hierarchical characteristics of spacecraft information, resulting in a large number of physically infeasible or logically contradictory invalid solutions during path search, thus limiting the accuracy of reasoning; on the other hand, it lacks sufficient modeling of the temporal dependencies and constraints of spacecraft behavior, making it difficult to quickly output reliable results in scenarios with stringent real-time requirements. Furthermore, with the surge in the number of spacecraft in orbit and the diversification of their behavior patterns, general multi-hop reasoning methods, lacking domain knowledge guidance, are prone to combinatorial explosion problems, and their computational efficiency cannot meet the high-efficiency requirements of aerospace missions.

[0004] Against this backdrop, optimizing knowledge graph multi-hop reasoning methods for the characteristics of the aerospace field has become crucial. By optimizing semantic modeling mechanisms, strengthening domain constraint embedding, and improving the efficiency of complex relationship processing, it is possible to accurately capture the deep logical connections of spacecraft behavior, solve the shortcomings of traditional methods in terms of accuracy, real-time performance, and reliability, and provide core technical support for the safe and efficient execution of aerospace missions. Summary of the Invention

[0005] The purpose of this invention is to provide an optimization method for spacecraft intent reasoning based on multi-hop knowledge chains. This method can significantly improve the accuracy, real-time performance, and engineering feasibility of spacecraft behavioral intent reasoning in complex scenarios.

[0006] To achieve the above objectives, the present invention provides the following solution: An optimization method for spacecraft intent reasoning based on multi-hop knowledge chains includes the following steps: S1. Establish a hierarchical knowledge framework for spacecraft behavior intention reasoning: define semantic associations step by step from top-level spacecraft entity information to bottom-level possible intentions, and clarify the logical relationship between subsystems, components, behaviors and intentions through visualization tools, and construct a dynamic subgraph that integrates spatiotemporal semantics. S2. Design a time-coding mechanism for orbital period perception: Convert event timestamps into periodic vector representations to capture the cyclical patterns of spacecraft orbital motion; integrate orbital dynamic parameters and use adaptive weighting to enable the encoding to generalize to different orbital types; S3. Construct a rule-guided hierarchical reasoning architecture: The first layer performs behavior pattern matching and track condition verification based on a predefined domain rule base, quickly filtering high-confidence intent candidate sets, and the rule confidence is dynamically adjusted according to historical trigger results; The second layer introduces a graph neural network, which aggregates node semantics, track and time information through multi-round feature propagation to generate a fine-grained intent probability distribution; S4. Perform temporal consistency verification of intent: Through dynamic time warping algorithm, the feature embedding of the current behavior sequence is compared with the historical intent template library to filter out abnormal intents that deviate significantly from the pattern; at the same time, based on predefined temporal logic constraint rules, a directed dependency graph of behavior sequence is constructed to automatically detect contradictions of circular dependency or temporal reversal, and intents that violate the core physical order are rejected. S5. Evolution of hierarchical inference architecture based on closed-loop feedback: By establishing a continuous optimization closed loop, at the model level, a training sample set is constructed using on-orbit inference error cases, and the neural network parameters are dynamically updated through regularized loss function and gradient optimization strategy; at the rule level, a periodic evaluation mechanism is established to automatically eliminate redundant rules with low confidence and long-term non-triggering, while high-value association rules are extracted from historical behavior by combining expert experience and data mining algorithms, and added to the rule base after feasibility verification.

[0007] Preferably, S1 further includes: Simultaneously, it embeds aerospace-specific constraint rules to ensure the autonomy of the graph; each on-orbit event is structured as a data unit containing a unique identifier, precise timestamp, orbital parameters, entity association, and intent attributes; based on a preset duration window, the event sequence is slidably truncated to dynamically generate a subgraph structure: the subgraph nodes cover all levels of entities within the window, and the edge relationships are divided into two categories: temporally dependent edges are constructed for events with adjacent times, and inherent edges are constructed based on the inherent relationships of the knowledge graph, forming a dynamic subgraph that integrates temporality and semantics.

[0008] Preferably, in S2, the event timestamps are converted into periodic vector representations to capture the cyclical patterns of the spacecraft's orbital motion, specifically including: The event timestamp t is mapped to the relative phase within the orbital period, calculated as follows:

[0009] in, This represents the remainder of the timestamp within the current orbital period. For the spacecraft's orbital period, the relative phase is encoded using sine and cosine functions to obtain a two-dimensional time embedding vector:

[0010] This encoding method utilizes the periodicity of trigonometric functions to transform linear time into a ring-shaped phase space, so that events spaced at integer intervals have the same encoding vector, thereby capturing the periodic patterns of spacecraft orbital motion.

[0011] Preferably, in S2, the orbital dynamics parameters are fused, and the encoding is made generalizable to different orbital types through adaptive weighting, specifically including: First, consider the six numbers of the orbit. After normalization, the six orbital elements include the semi-major axis. eccentricity Track inclination Right ascension of ascending node Perigeal argument True near point angle The parameters are scaled to a linear scale. The interval is denoted as Then, through the learnable parameter matrix and bias The normalized orbital parameters are coupled with time embedding:

[0012] in, To incorporate the temporal embedding vector with temporal encoding weights, It is the Sigmoid activation function. This represents matrix element multiplication; the model dynamically adjusts the weights of the time encoding based on the orbital eccentricity and inclination.

[0013] Preferably, in S3, the second layer introduces a graph neural network, which aggregates node semantics, trajectory, and time information through multiple rounds of feature propagation to generate a fine-grained intent probability distribution, specifically including: This paper introduces graph neural network semantic enhancement technology, which aggregates multimodal features through a lightweight GCN to generate intent representations containing temporal, orbital, and semantic information. Firstly, in the GCN network architecture, 3D one-hot vectors are used for encoding to distinguish node levels. This is automatically generated through a one-hot encoding layer. Simultaneously, orbital parameters are normalized, and the six orbital roots are... To standardize, the formula is:

[0014] in and Here are the historical mean and standard deviation of each parameter in orbit. The orbital six-root vector is stored in the parameter calibration table; the temporal embedding directly uses the 2D orbital period-aware encoding output from the input layer. After parameter preprocessing, graph convolution operation is performed: first, adjacency matrix processing is performed on the dynamic subgraph. Constructing an adjacency matrix Add a self-loop on the diagonal ,in, Given an N-order identity matrix, calculate the angle matrix. Subsequently, feature propagation is performed, using a two-layer GCN mapping to 64 dimensions and a second layer to 32 dimensions. The activation function is ReLU, and the feature concatenation uses the GCNConv layer from the PyTorchGeometric library. Finally, possible intent score prediction is performed, first by mean pooling, and then by applying mean pooling to the node feature matrix output by the GCN. Calculate the mean to obtain the subgraph representation. It uses two fully connected layers and finally a Softmax function probability distribution to characterize the confidence of possible intentions.

[0015] Preferably, S3 also includes: The multi-hop inference process employs directed acyclic graph modeling for path search, synchronously embeds orbital dynamics physical constraints, calculates the cumulative velocity increment of the path in real time, and performs immediate pruning on invalid paths that exceed the spacecraft's capabilities to ensure that the inference results meet engineering feasibility requirements.

[0016] Preferably, S4 also includes: Novel behavioral association patterns identified during the verification process are transformed into new temporal constraint rules after reliability assessment, thereby enabling the self-evolution of the knowledge chain.

[0017] Preferably, in S5, a training sample set is constructed at the model level using on-orbit inference error cases, and the neural network parameters are dynamically updated through a regularized loss function and gradient optimization strategy, specifically including: Introducing online learning and updating techniques, the GCN model parameters are dynamically optimized using in-orbit inference error cases. When the expected behavior corresponding to the intent does not occur, error cases are added to the training set, and event sequences, dynamic subgraph structures, and true intent labels are recorded simultaneously to form triples. ,in, For dynamic subgraphs, The feature matrix of the subgraph nodes. The vector represents the true intent label; the mean squared error with L2 regularization is used as the loss function.

[0018] Where N is the batch size and λ is the regularization coefficient. Including the weight matrix of GCN and fully connected layer parameters, Let i be the true intent label vector of the i-th sample. Let be the probability distribution vector of the model's predicted intent for the i-th sample; in terms of optimizer configuration, the Adam optimizer is used with an initial learning rate of 0.001, and the learning rate is adjusted exponentially every 1000 iterations.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a spacecraft intent reasoning optimization method based on a multi-hop knowledge chain as described above.

[0020] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention constructs a hierarchical knowledge chain of "spacecraft information → subsystems → components → behavior → possible intentions," integrating orbital dynamics constraints and temporal behavior modeling to generate dynamic subgraphs that incorporate spatiotemporal semantics. It employs a hierarchical reasoning architecture that combines rule-guided and graph neural network collaboration. At the rule layer, high-confidence intention candidates are rapidly screened and real-time pruning is performed using orbital velocity increments. At the GNN layer, multi-hop features are aggregated along the knowledge chain to generate a fine-grained intention distribution. A dual verification mechanism combining dynamic time warping with historical templates and predefined temporal logic constraints filters out physical contradictions and abnormal intentions. Furthermore, a closed-loop optimization is formed through error-case-driven online learning and dynamic evolution of the rule base, significantly improving the accuracy, real-time performance (reasoning latency reduced to milliseconds), and engineering feasibility of spacecraft intention reasoning in complex scenarios, providing core technical support for the safe and efficient execution of space missions. Attached Figure Description

[0021] 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.

[0022] Figure 1 A flowchart illustrating an optimization method for spacecraft intent reasoning based on multi-hop knowledge chains provided by this invention; Figure 2 This is a schematic diagram illustrating the construction process of the time feature vector of the present invention; Figure 3 This is a schematic diagram of the spacecraft intent reasoning and representation process of the present invention. Detailed Implementation

[0023] 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 embodiments of the present invention, and not all embodiments. 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.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] like Figure 1 As shown, the present invention provides a spacecraft intent reasoning optimization method based on multi-hop knowledge chains, comprising the following steps: S1. Establish a hierarchical knowledge framework for spacecraft behavior intention reasoning: define semantic associations step by step from top-level spacecraft entity information to bottom-level possible intentions, and clarify the logical relationship between subsystems, components, behaviors and intentions through visualization tools, and construct a dynamic subgraph that integrates spatiotemporal semantics. S2. Design a time-coding mechanism for orbital period perception: Convert event timestamps into periodic vector representations to capture the cyclical patterns of spacecraft orbital motion; integrate orbital dynamic parameters and use adaptive weighting to enable the encoding to generalize to different orbital types; S3. Construct a rule-guided hierarchical reasoning architecture: The first layer performs behavior pattern matching and track condition verification based on a predefined domain rule base, quickly filtering high-confidence intent candidate sets, and the rule confidence is dynamically adjusted according to historical trigger results; The second layer introduces a graph neural network, which aggregates node semantics, track and time information through multi-round feature propagation to generate a fine-grained intent probability distribution; S4. Perform temporal consistency verification of intent: On the one hand, through dynamic time warping algorithm, the feature embedding of the current behavior sequence is compared with the historical intent template library to filter out abnormal intents that deviate significantly from the pattern; on the other hand, based on predefined temporal logic constraint rules, a directed dependency graph of behavior sequence is constructed to automatically detect contradictions of circular dependency or temporal reversal, and intents that violate the core physical order are rejected. S5. Evolution of hierarchical inference architecture based on closed-loop feedback: By establishing a continuous optimization closed loop, at the model level, a training sample set is constructed using on-orbit inference error cases, and the neural network parameters are dynamically updated through regularized loss function and gradient optimization strategy; at the rule level, a periodic evaluation mechanism is established to automatically eliminate redundant rules with low confidence and long-term non-triggering, while high-value association rules are extracted from historical behavior by combining expert experience and data mining algorithms, and added to the rule base after feasibility verification.

[0026] Specifically, the method of the present invention includes: Step 1: Construct a knowledge graph reasoning chain that conforms to the reasoning logic of space target behavior intent. First, define a multi-layered semantic framework to clearly express the hierarchical relationship of "spacecraft information → subsystem → component → behavior → possible intent". The top layer defines spacecraft entity information, including subclasses such as satellites and probes, as well as the spacecraft's orbital information and motion status information; the middle layer is divided by function into attitude control subsystems, energy subsystems, etc., and further refined to specific components (such as attitude control engines and solar panels); the bottom layer associates the behaviors it performs (such as ignition and rotation) with their corresponding intents (such as orbit maintenance and power optimization). The hierarchical relationship is visualized using a UML class diagram, and aerospace-specific constraints are added, such as "each component must belong to at least one subsystem" and "the prerequisite behavior for orbit maintenance intent must include orbit change or attitude adjustment," to ensure the self-consistency of the knowledge graph logic.

[0027] Add hierarchical labels (levels 0-4, corresponding to spacecraft information, subsystems, components, behaviors, and possible intentions, respectively) to each node to indicate its position in the inference chain. Represent each event as a quintuple. ,in: A unique identifier for the event; The timestamps for the events are in UTC time format, accurate to the millisecond level, with intervals of approximately 5 seconds, to distinguish the times when different events occurred; The orbital number is six (semi-major axis) eccentricity Track inclination Right ascension of ascending node Perigeal argument True near point angle (This information can be collected in real time by sensors on the spacecraft's attitude and orbit control system.) It is a set of entity associations, determined by the "component-behavior" associations defined in the spacecraft knowledge graph, such as "attitude control engine-ignition" and "solar panel-rotation". It is a set of possible intent attributes, including the intent target object, intent confidence, and other information.

[0028] Use a fixed duration window T (e.g., 10 minutes) to slide through the timestamped event sequence. The sliding step size can be set to... (e.g., 1 minute). Each subgraph Includes all event nodes within the window (types include spacecraft information nodes, subsystem nodes, component nodes, behavior nodes, and potential intent nodes), as well as two types of edges: (1) Temporally dependent edges, when the time interval between two events is less than In this case, a temporal dependency edge is added between the two event nodes to characterize the possible correlation between the two events at the temporal level; (2) Knowledge graph inherent edges: Based on the predefined relationships such as “component → behavior” and “behavior → possible intention” in the knowledge graph, knowledge graph inherent edges are added between the corresponding nodes to represent the historical information correlation between the nodes at two levels.

[0029] Step 2, perform temporal feature encoding, such as... Figures 2-3 As shown, the core objective of time-position coding is to convert event timestamps into vector representations that contain the periodicity of the orbit. First, the orbital period is calculated. According to Kepler's third law, the orbital period of a spacecraft... From the semi-major axis The only certainty is that the calculation formula is:

[0030] in, For the gravitational parameters of the central celestial body, Mass of the central celestial body, semi-major axis of the orbit This formula characterizes the size of the orbital ellipse. Based on this formula, a quantitative relationship between orbital geometric parameters and the orbital period is established, providing a physical basis for periodic modeling of time characteristics.

[0031] The event timestamp t is mapped to the relative phase within the orbital period, calculated as follows:

[0032] in This represents the remainder of the timestamp within the current orbital period. The phase is encoded using sine and cosine functions to obtain a two-dimensional time embedding vector:

[0033] This encoding method utilizes the periodicity of trigonometric functions to transform linear time into a ring-shaped phase space, so that events spaced at integer intervals have the same encoding vector, thereby capturing the periodic patterns of spacecraft orbital motion (such as attitude adjustment, payload operation, and other behaviors that are often synchronized with the orbital period).

[0034] To further integrate orbital characteristics, orbital parameters are modulated through time embedding: First, the six orbital roots are... Normalization is performed, and the parameters are scaled to a normal value using a linear transformation. The interval is denoted as Then, through the learnable parameter matrix and bias The normalized orbital parameters are coupled with time embedding:

[0035] in It is the Sigmoid activation function. This represents matrix element multiplication. The model dynamically adjusts the weights of the time encoding based on parameters such as orbital eccentricity and inclination, enhancing its adaptability to different orbital types such as elliptical and inclined orbits.

[0036] Dynamic subgraph embedding can integrate the features of multiple node types within a sliding window into a unified subgraph representation, with nodes using one-hot encoding. The dimension of each component node is equal to the total number of component entities in the spacecraft knowledge graph. If the knowledge graph defines L component entities, for example, the encoding of the "solar panel" node is a vector of length L, with the corresponding position set to 1 and the rest set to 0; similarly, the dimension of each behavior node is equal to the total number of predefined behavior types. For example, the "attitude control motor ignition" node corresponds to a unique dimension, and the encoding vector is set to 1 at the position of that dimension and 0 at the other positions; the intent node is initialized to a zero vector. ,in The total number of intent types is preset. The features of intent nodes are dynamically generated during the inference phase using a graph convolutional network and initialized as zero vectors to avoid prior bias.

[0037] The initial features of each node are composed of three parts:

[0038] in, One-hot encoding for node type (component or behavior node) or zero vector (intent node); The orbital six-root number associated with the event for this node (only included for behavior nodes; for component and intent nodes, set to the default value under the spacecraft operational orbit type). The timestamps of the behavior nodes are set (the timestamps in component and intent nodes are set to 0 and are not included in the encoding). Node type information, orbital state parameters, and time periodic features are deeply fused through feature concatenation to form an initial node representation containing rich context.

[0039] An attention mechanism is used to aggregate the features of all nodes within the window to generate a subgraph. Initial representation: First, calculate the node importance weights. :

[0040] in For the set of nodes in the subgraph, , and For learnable parameters, Let be the hidden layer dimension of the attention mechanism. Then, a subgraph representation is obtained through weighted summation:

[0041] This process adaptively focuses on nodes more important to intent reasoning (such as frequently occurring key components or core behaviors) and suppresses interference from secondary nodes. In temporal context modeling, to capture the temporal dependencies between consecutive subgraphs, gated cyclic units are used to process the subgraph sequences generated by the sliding window.

[0042] in, The subgraph represents the hidden state of the previous time window. Through a recursive update mechanism, the current subgraph representation will include the temporal context information of the historical window, providing temporally coherent input features for subsequent multi-hop inference. A schematic diagram of the feature temporal encoding and feature fusion representation process based on orbital periodicity awareness is attached. Figure 3 As shown.

[0043] Step 3: Construct a rule-guided hierarchical reasoning architecture. Multi-hop intent reasoning is based on the collaborative design of "rule-guided hierarchical reasoning" and "orbital constraint path propagation," achieving efficient reasoning from spacecraft behavior sequences to possible intents. The rule-guided hierarchical reasoning architecture employs a hierarchical mechanism of "explicit rule fast filtering + graph neural network semantic enhancement" to achieve coarse screening and fine representation of intents, specifically including two technical sub-layers (illustrated below): (1) Rule filtering generates intent candidate technology. The goal is to build a rule base based on domain expert knowledge, quickly generate high-confidence intent candidates through logical condition matching, and narrow the scope of reasoning. The rule base is stored in a structured data format (such as JSON). The fields included in each rule given in the example are as follows; { / / A unique identifier for each rule, used to distinguish different rules (primary key) "rule_id": "R001", / / Behavioral pattern matching criteria: A list of consecutively occurring behavioral types. / / Example: Detect whether the "attitude control engine ignition" behavior occurs 3 times consecutively (array length determines the number of consecutive occurrences) "behavior_pattern": ["Attitude Control Engine Ignition", "Attitude Control Engine Ignition", "Attitude Control Engine Ignition"], / / Orbital parameter constraints (key-value pair structure) "orbit_condition": { / / Name of the track parameters to be verified (values ​​are the six track elements: i = inclination angle, a = semi-major axis, etc.) "parameter": "i", / / Comparison operators (supports >, <, =, etc., here meaning "greater than") "operator": ">", / / Threshold (unit corresponds to parameter, here "i" is angle, unit: degree) "threshold": 0.5, / / Threshold for the time interval between adjacent events (in seconds) / / Example: Require that the time difference between adjacent behavioral events be less than 60 seconds (1 minute). "time_window": 60 }, / / The type of intent generated after the rule is triggered (must be consistent with the predefined intent list) "intention": "track maintenance", / / Initial confidence level of the rule (value range: 0~1, reflecting the prior reliability of the rule) "initial_confidence": 0.8, / / Timestamp of the last time the rule was triggered (used for dynamic eviction strategies) / / Format: ISO 8601 standard time (YYYY-MM-DD HH:mm:ss) "last_trigger_time": "2025-05-29 12:00:00" } (2) Behavior pattern matching: The behavior type of K consecutive events (e.g., K=3) is detected by a sliding window to see if they completely match the preset rule pattern. The window sliding step is 1 event. Track parameters are also checked, and the track inclination angle of adjacent events is extracted. , The calculation checks whether the absolute difference exceeds 0.5° and whether the time interval between the two events is less than 60 seconds (1 minute) to determine whether the tilt angle has changed, and then to determine whether the attitude control engine ignition has occurred.

[0044] To enable dynamic updates to rule confidence, a separate rule manager is designed to periodically (e.g., after each rule is triggered) update the confidence based on historical data: statistically analyze historical rule trigger records and calculate the correct trigger rate. Incremental update formula adopted ,in To optimize the learning rate and avoid drastic fluctuations in confidence, when a rule is triggered, the trigger time and result (correct or incorrect) are recorded and stored in a historical log table for long-term analysis.

[0045] A graph neural network (GCN) semantic enhancement technique is introduced, aggregating multimodal features through a lightweight GCN to generate an intent representation containing temporal, orbital, and semantic information. Firstly, in the GCN network architecture, to distinguish node levels, 3D one-hot vectors are used for encoding (components are...). , behavior The possible intention is to It is automatically generated through a one-hot encoding layer, while the orbital parameters are normalized and the six orbital roots are... To standardize, the formula is:

[0046] in and The historical mean and standard deviation of each parameter in orbit are stored in the parameter calibration table; the temporal embedding directly uses the 2D orbital period sensing code output from the input layer. After parameter preprocessing, graph convolution operation is performed: first, adjacency matrix processing is performed on the dynamic subgraph. Constructing an adjacency matrix Add a self-loop on the diagonal And calculate the angle matrix. Subsequently, feature propagation is performed using a two-layer GCN (two layers have been verified to be sufficient to balance accuracy and efficiency). The first layer maps node features from 11 dimensions (3+6+2) to 64 dimensions, and the second layer maps them to 32 dimensions. The activation function is ReLU, and the GCNConv layer from the PyTorch Geometric library is used for feature concatenation, which can implement the custom input feature concatenation logic in step 2. Finally, possible intent score prediction is performed. First, mean pooling is applied to the node feature matrix output by the GCN. Calculate the mean to obtain the subgraph representation. Through two fully connected layers (32→16→ Finally, the probability distribution using the Softmax function is used to characterize the confidence level of possible intentions.

[0047] Based on the aforementioned GCN network design, path constraint propagation is implemented, and orbital dynamics principles are embedded in the reasoning process. Physical constraints are used to eliminate infeasible paths, ensuring the engineering feasibility of the intended action. Quantitative constraints on orbital change behavior and velocity increments are established based on orbital dynamics constraint modeling to avoid generating intentions with super-engine capabilities; these intentions are only used for reasoning related to orbital change behavior.

[0048] Simultaneously, path pruning strategies and multi-hop search techniques are introduced to verify constraints in real time during the inference path and efficiently prune invalid branches. The multi-hop search engine design path is represented using a directed acyclic graph (DAG), where nodes are (event, intent) pairs, and edges represent "behavior → intent" or "intent → behavior" relationships. The search algorithm uses breadth-first search, with a maximum search depth of 5 hops (an experimental value to avoid computational explosion). Each element in the queue contains a list of current path nodes and a cumulative speed increment. Subsequently, real-time verification is performed. When the path contains "track change" type behaviors (determined by the behavior node type identifier), the corresponding values ​​for that behavior are accumulated. arrive If the speed increment exceeds the spacecraft's maximum permissible speed increment, it exceeds actual physical constraints and should be pruned immediately. During the pruning rate monitoring and optimization process, the number of pruned paths in each inference iteration is recorded. Total number of paths And calculate the pruning rate. When the pruning rate exceeds 80%, it indicates that the constraints are too strict, and the GCN parameters will be automatically fine-tuned to balance physical constraints and inference coverage.

[0049] Step 4: Verify the temporal consistency of the intent. First, a dual mechanism of Dynamic Time Warping (DTW) matching and temporal logic verification is used to ensure the consistency between the inferred intent and historical behavioral patterns and physical order. Temporal alignment matching of behavioral sequences is required to filter out abnormal intents caused by occasional noise. During the ground testing phase, at least 100 typical behavioral sequences are collected for each preset intent, and corresponding embedding sequences are generated using GCN. ( For sequence length, (For GCN output dimensions), establish a structured template library, with the storage format as follows: ,in For the first Each historical template sequence is automatically added to the corresponding template library for each intent sequence that passes on-orbit verification.

[0050] The temporal logic of the inferred possible intentions is validated, and intentions that violate physical logic are excluded based on predefined event order constraints. The constraint rules are defined and stored in an XML file, as shown in the example below. <constraint> <precondition> Behavior ID="B001" (Solar panel turned on)< / precondition> <postcondition> Behavior ID="B002" (Solar panel rotation)< / postcondition> <description> Solar panels must be turned on before they rotate.< / description> < / constraint> Constraint types include mandatory constraints (such as "solar panel rotation" must occur after "solar panel opening") and priority constraints (such as "antenna rotation" must occur after "load rotation"). Violation of the former directly rejects the intent in the relevant reasoning chain, while the latter is only used as a scoring factor in intent reasoning.

[0051] Construct a directed acyclic graph (DAG), where nodes represent all event nodes in the current action sequence, labeled with event ID, type, and timestamp; edges are directed edges added from the "previous action" node to the "subsequent action" node according to constraint rules, and the timestamps must satisfy certain conditions. .

[0052] Predefined "pre- and post-" behavioral constraints are transformed into directed edges in a Directed Acyclic Graph (DAG), making abstract rules a computable graph structure. When the inference chain generates candidate intentions, DAG verification ensures that the corresponding behavioral sequence satisfies all explicit constraints. For example, if the "orbit holding" intention involves orbital change behavior, its associated "attitude control engine ignition → minor orbital change" must have a valid directed edge with an increasing timestamp; otherwise, the intention will be directly excluded. The trigger condition for inference knowledge chain correction is when the DAG detects a temporal contradiction, i.e., the existence of a cycle, a time-reversed edge, or an excessive DTW matching distance. When: If a DAG corresponding to a certain intention has a cycle (such as "attitude adjustment → orbit measurement" and "orbit measurement → attitude adjustment") forming a circular dependency, the reasoning path is immediately removed to avoid generating contradictory intentions; if a certain type of time reversal edge is frequently detected (such as "inter-satellite link disconnection" continuously appearing before "data transmission"), the system automatically marks the behavior pair as a potential new constraint, and determines whether to generate a new temporal constraint rule (such as "data transmission → inter-satellite link disconnection") through association rule mining, and adds it to the rule base after expert review to realize the dynamic evolution of constraint rules; when DTW matching fails and DAG verification passes (i.e., the behavior pattern is abnormal but the logic is legal), the case is added to the training set as a "boundary sample", and the temporal embedding weight of the GCN model is adjusted through online learning, such as enhancing the sensitivity of timestamp encoding to abnormal intervals to avoid misjudgment caused by fluctuations in the interval of a single behavior.

[0053] Step 5: Establish a closed-loop feedback mechanism to achieve system self-optimization and long-term adaptation through online learning and rule base evolution. Introduce online learning and update technology, utilizing on-orbit verification of error cases to dynamically optimize GCN model parameters. When the expected behavior corresponding to an intent fails to occur (e.g., the possible intent is "track hold" but the inclination continues to drift), it is automatically marked as a potential error case. When adding error cases to the training set, the event sequence, dynamic subgraph structure, and true intent labels are recorded simultaneously to form triples. ,in, It is a dynamic subgraph constructed from the sequence of spacecraft on-orbit events, inherent connections of the knowledge graph, and temporal dependency edges, used to capture the relationship structure between spacecraft behavior and components; This is the feature matrix of the subgraph nodes, which contains the feature representations of each node (component node, behavior node, intent node) in the subgraph; This is the true intent label vector (one-hot encoded form), used to label the true spacecraft intent corresponding to this error case. It serves as the benchmark for comparing the predicted value with the true value in the loss function calculation; and the mean squared error with L2 regularization is used as the loss function. Where the batch size N=32, and the regularization coefficient λ=0.01, Including the weight matrix of GCN Regarding the fully connected layer parameters and optimizer configuration, the Adam optimizer is used with an initial learning rate of 0.001. The learning rate is adjusted exponentially (decay rate 0.95) every 1000 iterations, and the gradient pruning threshold is set to 5. A dynamic rule base evolution technique is introduced to continuously optimize the explicit rule base, eliminating invalid rules and introducing new ones. The main strategy for rule elimination is periodic scanning, with a set time interval (e.g., 6 months). During this period, rule base cleanup is performed, and the rules being checked are: confidence levels. (This indicates a correct trigger rate of <50% within this period), and the last trigger time is more than 1 year ago (indicating a rule that has not been used for a long time). For rules that meet the conditions, they are first marked as "to be phased out". After a 3-month buffer period, if they are still not triggered, they are permanently deleted to avoid accidentally deleting occasionally effective rules. At the same time, new rules are generated based on expert import and data mining: through the human-computer interaction interface of the ground control center, experts can manually input rules. The format is compatible with the existing rule base and automatic syntax verification is performed. The Apriori algorithm can also be used to mine association rules from historical event sequences. The preprocessed event sequence is used as input, and the minimum support is set to 0.1 (indicating that at least 10% of the sequence contains the antecedent of the rule), the minimum confidence is 0.7, and the maximum rule length is 3 to avoid overly complex rules. The mined rules are verified for physical feasibility through the ground simulation system, and after manual review, they are added to the rule base to realize the dynamic updating of the intent reasoning rule base.

[0054] The invention also includes a spacecraft intent reasoning optimization system based on multi-hop knowledge chains, comprising three modules: a temporal behavior modeling and dynamic subgraph extraction module; a knowledge chain-based multi-hop intent reasoning module; and an intent verification and dynamic correction module for reasoning rules. The temporal behavior modeling and dynamic subgraph extraction module constructs a spatiotemporal representation system for spacecraft behavior. A hierarchical knowledge framework is established, defining the semantic chain from spacecraft entities to underlying intentions, ensuring logical consistency through domain constraints. Original events are parsed into structured quintuples, integrating timestamps, orbital parameters, and entity associations. Based on a sliding window mechanism, event sequences are dynamically extracted to generate subgraph structures that combine temporal dependency edges with inherent edges of the knowledge graph.

[0055] The knowledge chain-based intent multi-hop reasoning module enables hierarchical reasoning through the collaboration of rules and neural networks. The upper layer employs an explicit rule engine that predefines combinations of behavioral patterns and physical constraints on the track to quickly filter and generate high-confidence intent candidates, with rule confidence dynamically evolving based on the trigger result. The lower layer deploys a lightweight graph neural network that propagates node features along the knowledge chain hierarchy, aggregates semantic information through multi-round graph convolution, and outputs the intent probability distribution.

[0056] The intent verification and inference rule dynamic correction module verifies the reliability of intent inference through dynamic temporal matching and logical dependency verification technology, thereby filtering out abnormal results and eliminating physical contradictions. At the same time, it dynamically optimizes the inference model based on the error case-driven online learning mechanism. Combined with the rule base dynamic evolution technology, the system can achieve self-correction and continuous adaptation to ensure the accuracy of long-term inference.

[0057] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the spacecraft intent reasoning optimization method based on a multi-hop knowledge chain as described above.

[0058] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0059] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A spacecraft intent reasoning optimization method based on multi-hop knowledge chains, characterized in that, Includes the following steps: S1. Establish a hierarchical knowledge framework for spacecraft behavior intention reasoning: define semantic associations step by step from top-level spacecraft entity information to bottom-level possible intentions, and clarify the logical relationship between subsystems, components, behaviors and intentions through visualization tools, and construct a dynamic subgraph that integrates spatiotemporal semantics. S2. Design a time-coding mechanism for orbital period perception: Convert event timestamps into periodic vector representations to capture the cyclical patterns of spacecraft orbital motion; integrate orbital dynamic parameters and use adaptive weighting to enable the encoding to generalize to different orbital types; S3. Construct a rule-guided hierarchical reasoning architecture: The first layer performs behavior pattern matching and track condition verification based on a predefined domain rule base, quickly filtering high-confidence intent candidate sets, and the rule confidence is dynamically adjusted according to historical trigger results; The second layer introduces a graph neural network, which aggregates node semantics, trajectory and time information through multi-round feature propagation to generate a fine-grained intent probability distribution; S4. Perform temporal consistency verification of intent: Through dynamic time warping algorithm, the feature embedding of the current behavior sequence is compared with the historical intent template library to filter out abnormal intents that deviate significantly from the pattern; at the same time, based on predefined temporal logic constraint rules, a directed dependency graph of behavior sequence is constructed to automatically detect contradictions of circular dependency or temporal reversal, and intents that violate the core physical order are rejected. S5. Evolution of hierarchical inference architecture based on closed-loop feedback: By establishing a continuous optimization closed loop, at the model level, a training sample set is constructed using on-orbit inference error cases, and the neural network parameters are dynamically updated through regularized loss function and gradient optimization strategy; at the rule level, a periodic evaluation mechanism is established to automatically eliminate redundant rules with low confidence and long-term non-triggering, while high-value association rules are extracted from historical behavior by combining expert experience and data mining algorithms, and added to the rule base after feasibility verification.

2. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 1, characterized in that, S1 also includes: Simultaneously, it embeds aerospace-specific constraint rules to ensure the autonomy of the graph; each on-orbit event is structured as a data unit containing a unique identifier, precise timestamp, orbital parameters, entity association, and intent attributes; based on a preset duration window, the event sequence is slidably truncated to dynamically generate a subgraph structure: the subgraph nodes cover all levels of entities within the window, and the edge relationships are divided into two categories: temporally dependent edges are constructed for events with adjacent times, and inherent edges are constructed based on the inherent relationships of the knowledge graph, forming a dynamic subgraph that integrates temporality and semantics.

3. The spacecraft intent reasoning optimization method based on multi-hop knowledge chains according to claim 1, characterized in that, In step S2, the event timestamps are converted into periodic vector representations to capture the cyclical patterns of spacecraft orbital motion, specifically including: The event timestamp t is mapped to the relative phase within the orbital period, calculated as follows: in, This represents the remainder of the timestamp within the current orbital period. For the spacecraft's orbital period, the relative phase is encoded using sine and cosine functions to obtain a two-dimensional time embedding vector: This encoding method utilizes the periodicity of trigonometric functions to transform linear time into a ring-shaped phase space, so that events spaced at integer intervals have the same encoding vector, thereby capturing the periodic patterns of spacecraft orbital motion.

4. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 3, characterized in that, In S2, the orbital dynamics parameters are fused, and adaptive weighting is used to enable the encoding to generalize to different orbital types. Specifically, this includes: First, consider the six numbers of the orbit. After normalization, the six orbital elements include the semi-major axis. eccentricity Track inclination Right ascension of ascending node Perigeal argument True near point angle The parameters are scaled to a linear scale. The interval is denoted as Then, through the learnable parameter matrix and bias The normalized orbital parameters are coupled with time embedding: in, To incorporate the temporal embedding vector with temporal encoding weights, It is the Sigmoid activation function. This represents matrix element multiplication; the model dynamically adjusts the weights of the time encoding based on the orbital eccentricity and inclination.

5. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 1, characterized in that, In S3, the second layer introduces a graph neural network, which aggregates node semantics, trajectory, and time information through multi-round feature propagation to generate a fine-grained intent probability distribution, specifically including: This paper introduces graph neural network semantic enhancement technology, which aggregates multimodal features through a lightweight GCN to generate intent representations containing temporal, orbital, and semantic information. Firstly, in the GCN network architecture, 3D one-hot vectors are used for encoding to distinguish node levels. This is automatically generated through a one-hot encoding layer. Simultaneously, orbital parameters are normalized, and the six orbital roots are... To standardize, the formula is: in, and This is a vector of the historical mean and standard deviation of each parameter of the orbital six-root system. The orbital six-root vector is stored in the parameter calibration table; the temporal embedding directly uses the 2D orbital period-aware encoding output from the input layer. After parameter preprocessing, graph convolution operation is performed: first, adjacency matrix processing is performed on the dynamic subgraph. Constructing an adjacency matrix Add a self-loop on the diagonal ,in, Given an N-order identity matrix, calculate the angle matrix. Subsequently, feature propagation is performed, using a two-layer GCN mapping to 64 dimensions and a second layer to 32 dimensions. The activation function is ReLU, and the feature concatenation uses the GCNConv layer from the PyTorch Geometric library. Finally, possible intent score prediction is performed, first by mean pooling, and then by applying mean pooling to the node feature matrix output by the GCN. Calculate the mean to obtain the subgraph representation. It uses two fully connected layers and finally a Softmax function probability distribution to characterize the confidence of possible intentions.

6. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 5, characterized in that, S3 also includes: The multi-hop inference process employs directed acyclic graph modeling for path search, synchronously embeds orbital dynamics physical constraints, calculates the cumulative velocity increment of the path in real time, and performs immediate pruning on invalid paths that exceed the spacecraft's capabilities to ensure that the inference results meet engineering feasibility requirements.

7. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 1, characterized in that, S4 also includes: Novel behavioral association patterns identified during the verification process are transformed into new temporal constraint rules after reliability assessment, thereby enabling the self-evolution of the knowledge chain.

8. The spacecraft intent reasoning optimization method based on multi-hop knowledge chain according to claim 5, characterized in that, In S5, at the model level, a training sample set is constructed using on-orbit inference error cases, and the neural network parameters are dynamically updated through a regularized loss function and gradient optimization strategy. Specifically, this includes: Introducing online learning and updating techniques, the GCN model parameters are dynamically optimized using in-orbit inference error cases. When the expected behavior corresponding to the intent does not occur, error cases are added to the training set, and event sequences, dynamic subgraph structures, and true intent labels are recorded simultaneously to form triples. ,in, For dynamic subgraphs, The feature matrix of the subgraph nodes. The vector represents the true intent label; the mean squared error with L2 regularization is used as the loss function. Where N is the batch size and λ is the regularization coefficient. Including the weight matrix of GCN and fully connected layer parameters, Let i be the true intent label vector of the i-th sample. Let be the probability distribution vector of the model's predicted intent for the i-th sample; in terms of optimizer configuration, the Adam optimizer is used with an initial learning rate of 0.001, and the learning rate is adjusted exponentially every 1000 iterations.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a spacecraft intent reasoning optimization method based on a multi-hop knowledge chain as described in any one of claims 1 to 8.

Images