Spacecraft confrontation intention recognition method with interpretability

Abstract

The invention relates to a spacecraft confrontation intention recognition method with interpretability, which adopts a mode of combining deep learning and data mining to recognize the confrontation intention of a spacecraft and realize the tracing of the confrontation intention, and specifically comprises the following steps of: firstly, collecting time sequence state data of the spacecraft through a space-based or foundation remote sensing system; performing data processing on the original data, and performing association rule mining on historical confrontation data of the spacecraft through an association rule mining algorithm to obtain an association rule of state, action, moment-intention; meanwhile, a depth model is constructed and trained for intention recognition, and final spacecraft confrontation intention probability distribution is output. The result of the association rule mining algorithm is used for screening in adversarial intention recognition, so that the intention recognition precision is improved; and traceability of the recognition result of the confrontation intention is realized, the source of the confrontation intention is analyzed, an auxiliary decision is provided for a decision maker, and interpretability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of spacecraft anti-aggression intent recognition, and more particularly to an interpretable method for spacecraft anti-aggression intent recognition. Background Technology

[0002] With the development of artificial intelligence and big data technologies, spacecraft behavior and intent recognition has become a research hotspot in fields such as situational awareness and intelligent interaction. Against this backdrop, deep learning methods, with their automatic feature extraction capabilities, are widely applied to intent recognition tasks.

[0003] Currently, a great deal of scientific research and application has been carried out in the field of intent recognition, and fruitful results have been achieved. For example, Zhou Wangwang et al. proposed an intent recognition model based on deep neural networks in the literature (Zhou Wangwang. Airborne target combat intent recognition based on deep neural networks [J]. Acta Aeronautica Sinica, 2018, 39(11):200-208.). Under the condition of insufficient domain expert knowledge, the model obtains the rules between feature states and intents through self-training, effectively recognizing the combat intent of airborne targets. The patent (Wang Chuanyun, Zhang Yingying, Gao Qian, et al. An intent recognition method based on UAV swarm three-dimensional trajectory analysis [P]. 202411646538.X. Shenyang Aerospace University. 2025-03-04.) proposed an intent recognition method based on UAV swarm three-dimensional trajectory analysis. Combining deep learning networks, the model captures the temporal dependencies of three-dimensional trajectories, improving the ability to analyze complex behavioral patterns during intent recognition. A method and system for identifying the flight intention of unmanned aerial vehicles (UAVs) are proposed. A multi-dimensional behavior expectation model is constructed by combining a pre-set behavior pattern library and airspace usage rules to evaluate and predict the optimal flight intention.

[0004] However, most existing methods focus on improving the performance of a single model, often lacking a systematic exploration of the synergistic mechanism between deep features and rule-based knowledge. Especially in spacecraft trajectory scenarios, due to the increased complexity and interactivity of behavioral intentions, relying solely on deep models can easily lead to interpretability gaps, while relying solely on rule-based reasoning may suffer from insufficient expressive power. Therefore, a hybrid intention recognition method integrating deep representation learning and interpretable rule mining is needed, balancing accuracy and interpretability, and applicable to understanding spacecraft behavior in complex trajectory environments such as nonlinearity and weak supervision. Summary of the Invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide an interpretable method for identifying spacecraft adversarial intent, addressing the shortcomings of the prior art.

[0006] To address the aforementioned technical problems, this invention discloses an interpretable method for recognizing adversarial intentions in spacecraft.

[0007] An interpretable method for recognizing adversarial intent in spacecraft includes the following steps:

[0008] Step 1: Collect time-series status data of spacecraft through space-based or ground-based remote sensing systems, including the number of spacecraft, spacecraft ID number, attitude information (position, velocity, etc.), orbital information (low-Earth orbit, high-Earth orbit, etc.), fuel information data of each spacecraft, etc., and perform preprocessing on the raw time-series status data, including denoising, normalization and balancing, to construct a standardized adversarial sample set.

[0009] Step 2: Use the Apriori algorithm to mine association rules in the historical adversarial data of spacecraft, set minimum support and minimum confidence thresholds, and mine association rules for [state, action, time → intention].

[0010] Step 3: Construct and train an improved PBT-Net (Parallel BiGRU-Transformer Network) adversarial intent recognition model. Construct a deep neural network that includes a feature embedding layer, a multi-path parallel feature extraction layer, and a feature fusion layer. At the same time, use the mined association rule results [state, action, time → intent] to filter adversarial intents; where → indicates that the state, action, and time guide the intent.

[0011] Step 4: Output the results and probabilities of the adversarial intent, and at the same time, use association rules to trace the source of the results to obtain the path of adversarial intent identification, so as to provide a basis for commanders' decision-making and assist in decision-making.

[0012] The temporal state data mentioned in step 1 is represented as: a set of four-tuples: [state, action, intention, next moment's state]

[0013]

[0014] in, It represents spacecraft status data, including attitude information, orbital information, and fuel information; Indicates an action command. For specific intent, This is the state at the next moment.

[0015] Let the state vector of the principal spacecraft (i.e., the object of the identification intent) numbered B at time t be:

[0016]

[0017]

[0018] in, This represents the x-coordinate and y-coordinate of spacecraft B on a two-dimensional plane. This indicates its lateral and longitudinal velocities on the two-dimensional orbital plane. For remaining fuel resources, This is the load condition.

[0019] In step 1, the preprocessing specifically includes:

[0020] Step 1-1: Record the state of each spacecraft at time t from a simulation platform or remote sensing system at a fixed frequency (e.g., 10 Hz). Then, preprocess the sample set and calculate the mean of each time-series state data (including position, velocity, etc.). and standard deviation Eliminate those that meet the requirements Outlier samples

[0021] For a small number of missing measurements, linear interpolation is used to fill in the missing values:

[0022]

[0023] Steps 1-2: To address the class imbalance problem in the dataset, the SMOTE algorithm is introduced to oversample the minority class. Specifically, for each minority class sample... Randomly select a neighbor from its k nearest neighbors. Generate new synthetic samples:

[0024]

[0025] The formula indicates along and The straight line between them is interpolated at a random ratio to generate new samples. This process is repeated until the number of samples in each class is balanced. SMOTE oversampling avoids the overfitting risk of simply copying data, while improving the model's ability to recognize a few specific intentions.

[0026] Step 2 specifically involves processing the historical data, discretizing the time-series state data D to obtain discrete data Z, and then using the Apriori algorithm to mine frequent itemsets. And generate a set of association rules:

[0027]

[0028] in, , For minimum support, The minimum confidence level is represented by `supp`, which indicates the support level, and `conf` indicates the confidence level.

[0029] The feature embedding layer described in step 3 uses a one-dimensional convolutional neural network (1D-CNN) to extract local spatial features. Its input is preprocessed temporal state data. Multiple convolutional kernels of different widths are used in parallel to capture the local dependencies and patterns of features at different time scales. At the same time, during the model training process, the results of association rule mining [state, action, time → intent] are introduced in the form of constraints or auxiliary loss terms to perform logical filtering and consistency correction on the intent prediction output by the model, so as to improve the interpretability and accuracy of intent recognition.

[0030] The multi-parallel feature extraction layer is divided into parallel branch A and parallel branch B. Parallel branch A adopts an improved bidirectional gated recurrent unit (BiGRU) and introduces a temporal attention mechanism after the BiGRU layer to focus on features at key time steps and capture local temporal dynamic changes. Parallel branch B adopts an improved Transformer encoder and uses a multi-head self-attention mechanism for parallel computation. It retains the original encoder part of the Transformer and replaces the original decoder part with a prediction head. The Transformer can capture long-distance temporal dependencies and global contextual information.

[0031] In step 4, the final spacecraft adversarial intent and probability are output through a fully connected layer and a softmax classifier. At the same time, the set of association rules mined in step 2 is used to reverse trace the predicted adversarial intent to generate a traceable reasoning path. The path can provide intuitive and interpretable auxiliary decision-making information for the commander's situation assessment and decision-making.

[0032] The intent categories include reconnaissance, surveillance, guidance, interception, and escape.

[0033] The method is applicable to constrained game spaces, including but not limited to typical scenarios such as spatial formation, cluster task allocation, and intersection decision-making.

[0034] Beneficial effects:

[0035] (1) Strong temporal modeling capability: The parallel BiGRU and Transformer (PBT-Net) deep feature extraction module has both local spatial feature capture and long-short-term dependency modeling capabilities, and is suitable for deep expression of spacecraft dynamic behavior sequences.

[0036] (2) High precision and high robustness: By using SMOTE sample balancing technology and multi-layer deep neural network structure, the model’s recognition accuracy and generalization ability are enhanced in scenarios with imbalanced samples and complex state changes.

[0037] (3) Strong interpretability: The Apriori association rule algorithm is introduced to mine high-confidence behavioral rules from the original discretized features and deep features, thereby improving the interpretability and traceability of the classification results.

[0038] (4) Decision support mechanism: Association rule mining verifies, filters and traces the combat intention predictions output by the neural network, thereby providing decision support path for commanders.

[0039] (5) Fusion reasoning mechanism: By fusing association rules with deep networks, the advantages of rule-driven and data-driven approaches are combined, effectively improving the reliability and stability of intent recognition.

[0040] (6) Adaptable to complex environments: It is suitable for strategic game scenarios with high dynamics and high uncertainty, such as spacecraft orbital games and UAV swarm behavior analysis, and has good scalability and engineering application prospects. Attached Figure Description

[0041] Figure 1 Explainable Spacecraft Adversarial Intent Recognition Model Diagram

[0042] Figure 2 Explainable flowchart for spacecraft adversarial intent identification;

[0043] Figure 3 Schematic diagram of the improved BiGRU structure;

[0044] Figure 4 Improved Transformer structure diagram;

[0045] Figure 5 Schematic diagram of the improved BiGRU and Transformer parallel architecture. Detailed Implementation

[0046] This invention discloses an interpretable method for recognizing adversarial intent in spacecraft. First, it collects temporal state data of the spacecraft using a space-based or ground-based remote sensing system and processes the raw data (including denoising, normalization, and balancing). Second, it uses the Apriori association rule mining algorithm to mine association rules from the spacecraft's historical adversarial data, obtaining association rules for "state, action, time → intent". Simultaneously, it constructs and trains an improved PBT-Net deep model for intent recognition. This model introduces a parallel architecture with branches of an improved BiGRU and a Transformer. A temporal attention mechanism is introduced after the BiGRU layer to focus on features at key time steps and capture local temporal dynamic changes. In the Transformer layer, a multi-head self-attention mechanism is used for parallel computation. The original Encoder part of the Transformer is retained, and the original Decoder part is replaced with a prediction head. The Transformer can capture long-distance temporal dependencies and global contextual information. Finally, a fully connected layer and a softmax classifier are used to output the final probability distribution of spacecraft adversarial intentions (reconnaissance, surveillance, guidance, interception, escape, etc.). The results of association rule mining algorithms are used to filter adversarial intention identification, improving the accuracy of intention identification. In addition, it is possible to reverse trace the adversarial intention identification results, analyze the path of adversarial intentions, provide decision-making assistance to decision-makers, and enhance interpretability.

[0047] An interpretable method for recognizing adversarial intent in spacecraft includes the following steps:

[0048] Step 1: Collect time-series status data of spacecraft through space-based or ground-based remote sensing systems, including the number of spacecraft, spacecraft ID number, attitude information (position, velocity, etc.), orbital information (low-Earth orbit, high-Earth orbit, etc.), fuel information data of each spacecraft, etc., and perform preprocessing on the raw time-series status data, including denoising, normalization and balancing, to construct a standardized adversarial sample set.

[0049] Step 2: Use the Apriori algorithm to mine association rules in the historical adversarial data of spacecraft, set minimum support and minimum confidence thresholds, and mine association rules for [state, action, time → intention].

[0050] Step 3: Construct and train an improved PBT-Net (Parallel BiGRU-Transformer Network) adversarial intent recognition model. Construct a deep neural network that includes a feature embedding layer, a multi-path parallel feature extraction layer, and a feature fusion layer. At the same time, use the mined association rule results [state, action, time → intent] to filter adversarial intent.

[0051] Step 4: Output the results and probabilities of the adversarial intent, and at the same time, use association rules to trace the source of the results to obtain the path of adversarial intent identification, so as to provide a basis for commanders' decision-making and assist in decision-making.

[0052] 1) Collect the number of spacecraft and their ID numbers using space-based or ground-based remote sensing systems. Simultaneously, collect attitude information (position, velocity, etc.), orbital information (low Earth orbit, high Earth orbit, etc.), and fuel information for each spacecraft, and preprocess and balance the data. The specific steps are as follows:

[0053] Step 1: Record the state of 10 spacecraft at time t using a fixed frequency (e.g., 10 Hz) from the simulation platform. , , ..., .

[0054] Next, the sample set is preprocessed, and the mean of each dimension is calculated. and standard deviation Eliminate those that meet the requirements Outlier samples.

[0055] For a small number of missing measurements, linear interpolation is used to fill in the missing values:

[0056]

[0057] Step 2: To address the class imbalance problem in the dataset, the SMOTE algorithm is introduced to oversample the minority class. Specifically, for each minority class sample... Randomly select a neighbor from its k nearest neighbors. Generate new synthetic samples:

[0058]

[0059] The formula indicates along and The straight line between them is interpolated at a random ratio to generate new samples. This process is repeated until the number of samples in each class is balanced. SMOTE oversampling avoids the overfitting risk of simply copying data, while improving the model's ability to recognize a few specific intentions.

[0060] 2) The Apriori algorithm is used to mine association rules from historical spacecraft data to obtain the logic (rules) of the game-like competition between spacecraft, as well as the correlation correction between the rules and the model output. The specific steps are as follows:

[0061] Step 1: Construct a transaction set for association rule analysis using the preprocessed spacecraft state data. Treat the discretized state features of each sample as a term, and the corresponding intent category as a term in a transaction. For example, state features 1, 2, 3 (i.e., sequential state data) and the actual intent form a transaction Tj, where 1 ≤ j ≤ i. All transaction sets constitute the transaction database. ={T1, T2, ..., Ti}, the transaction set of the association rule can be represented as follows:

[0062] Table 1. Example of a transaction set for an association rule

[0063] Transaction ID State characteristic 1 (orbital change) State Feature 2 (Attitude Adjustment) State characteristic 3 (load activity) Actual Intent (Category) T1 Increase in semi-major axis of track Pitch angle adjustment Measurement and control signal enhancement orbit maintenance T2 Perimeter altitude decrease Frequent attitude maneuvers Radar powered on Space reconnaissance T3 Perimeter altitude decrease Frequent attitude maneuvers robotic arm unfolds Space docking T4 Perimeter altitude decrease Frequent attitude maneuvers Radar powered on Space reconnaissance T5 Increase in semi-major axis of track Pitch angle adjustment Measurement and control signal enhancement orbit maintenance … … … … … Ti … … … …

[0064] Step 2: Transaction database Execute the Apriori algorithm to find frequent itemsets between state features and intent labels. Define the support of any itemset A→B as:

[0065]

[0066] Where A represents features 1, 2, and 3 in the state features, B represents the actual intention, Num(*) represents the number of *, Num_total represents i (the total number of transactions) in the instance, and δ represents the support threshold. Through iterative incrementing frequency, itemsets with support not lower than the preset threshold are selected as frequent itemsets. .

[0067] Step 3: Generate association rules based on frequent itemsets. For each frequent itemset... The generation form is The rules ( , ).

[0068] Calculate the confidence level of the rule:

[0069]

[0070] Only retain those that meet the requirements. and The rules, where ε represents the support threshold, form the final intent association rule set. .

[0071] Step 4: Describe the discovered association rules, such as {decreasing perigee altitude, frequent attitude maneuvers} → {space reconnaissance}, {increased orbital semi-major axis, enhanced telemetry and control signals} → {orbit maintenance}, etc.

[0072] 3) Feature extraction is performed using the PBT-Net model to identify the spacecraft's countermeasures intent (reconnaissance, guidance, interception, etc.). The specific steps are as follows:

[0073] Step 1: Denoise and reconstruct the input time-series state data, and extract latent features using a 1D-CNN. The design is implemented in the following steps:

[0074] The preprocessed temporal data is input into a one-dimensional CNN network to extract local spatial features. Let the input sequence be... ,in Let be the feature vector at time j. For a convolution kernel of length k... bias The convolution output at position j is:

[0075]

[0076] in, It is a non-linear activation function (such as ReLU). By using multiple convolutional kernels and stacking multiple layers, it is possible to capture temporal features of different receptive fields.

[0077] The second step involves inputting the feature sequences extracted by the CNN into the improved BiGRU module to capture the forward and backward dependencies of the sequences; it also inputs them into the improved Transformer for feature extraction. The specific computation process of the improved BiGRU is as follows:

[0078]

[0079]

[0080]

[0081]

[0082] in, To update the door, To reset the door, This is a candidate hidden state. For the Sigmoid function, This is an element-wise product. BiGRU computes the product in parallel in both the forward and reverse directions on the same sequence. And the two are spliced ​​together to obtain the final representation. .

[0083] The specific calculation process of the improved Transformer is as follows:

[0084] For the input feature matrix Generate queries through linear mapping , key , value :

[0085]

[0086] The formula for calculating attention is:

[0087]

[0088] in, The dimension is the key vector dimension, used for scaling. A multi-head attention mechanism is employed.

[0089]

[0090]

[0091] Finally, the output features of the improved BiGRU and the output features of the Transformer are fused to obtain the fused feature layer:

[0092]

[0093] Where TotalF represents the result of feature layer fusion.

[0094] Step 4: The obtained feature layer TotalF is processed by a Softmax classifier to output the intent probability. Let the obtained feature vector be... Where C is the number of intent categories, and the predicted probability of the j-th category is obtained through the Softmax function:

[0095]

[0096] in, This is the linear mapping score for the corresponding category. The output is the deep model's prediction of the spacecraft's adversarial intent.

[0097] To improve overall recognition accuracy, SCDAE and MLP employ an end-to-end joint training strategy, optimizing the objective function as a weighted sum of the losses of the two modules:

[0098]

[0099] in, The reconstruction loss is for the autoencoder. Cross-entropy loss for intention classification; is a weighting coefficient used to adjust the relative importance between reconstruction and classification.

[0100] During training, the results of association rule mining are used to filter the adversarial intent recognition results. For example, the validation set in training has state features: the semi-major axis of the orbit increases and the telemetry and control signal is enhanced, and the output result is spatial reconnaissance. This result contradicts the rule set {the semi-major axis of the orbit increases and the telemetry and control signal is enhanced} → {orbit maintenance}, so it is deleted and not used as a training result.

[0101] 4) By fusing feature extraction and rule mining to determine intent, the high-precision classification capability of deep learning models is preserved, while providing clear and auditable rule-based support for decision-making, thus enhancing the interpretability of adversarial intent recognition. The specific implementation steps are as follows:

[0102] Step 1: For each spacecraft, extract its adversarial intent category probability distribution. Simultaneously, search for the set of association rules triggered by the spacecraft. .

[0103] The second step involves the commander predicting the spacecraft's adversarial intentions using a model. The model outputs the intentions and their probabilities, such as orbit maintenance (75%), space reconnaissance (10%), and space docking (8%). Association rule mining traces the path of these intentions. For example, if the output is orbit maintenance (75%), association rule mining will simultaneously provide {increased semi-major axis, significant pitch adjustment, enhanced telemetry and control signals} → {orbit maintenance}. This helps the commander clearly grasp real-time information and assists in subsequent decision-making.

[0104] This invention has a simple structure, strong adaptability, high decision-making efficiency, and also has the following characteristics:

[0105] (1) The generated adversarial intent is interpretable, and the adversarial intent is traced back to obtain the intent identification path, thereby providing auxiliary decision-making for the commander.

[0106] (2) It has strong generalization ability and is applicable to all spacecraft adversarial game scenarios.

[0107] (3) It has a fast recognition speed and high efficiency, which can better meet the user's needs for adversarial intent recognition.

[0108] (4) Simple to operate, easy to use, and highly stable.

[0109] Existing technologies (such as the deep neural network model developed by Zhou Wangwang et al. and UAV trajectory analysis methods) primarily focus on optimizing the performance of single models, failing to systematically address key technical bottlenecks in spacecraft adversarial intent recognition, including incomplete temporal feature modeling, failure to recognize a few intent categories, lack of basis for decision-making processes, and weak anti-interference capabilities of recognition results. Therefore, this invention provides an interpretable spacecraft adversarial intent recognition method. This method not only ensures high recognition accuracy through deep networks but also innovatively introduces an association rule mining mechanism, enabling a traceable reasoning path in the recognition process. Specifically, the system can verify, filter, and trace the intent predictions output by the neural network based on the mined association rules between state, action, time, and intent, thereby providing a decision support chain for commanders and significantly improving the interpretability, reliability, and engineering practical value of intent recognition in complex adversarial environments.

[0110] This invention provides an interpretable approach and method for identifying adversarial intentions in spacecraft. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A spacecraft adversarial intent recognition method with interpretability, characterized in that, Includes the following steps: Step 1: Collect time-series state data of spacecraft through space-based or ground-based remote sensing systems, and preprocess the raw time-series state data to construct a standardized set of adversarial examples; Step 2: Perform association rule mining on the historical adversarial data of spacecraft, set minimum support and minimum confidence thresholds, and mine association rules for [state, action, time → intention]. Step 3: Construct and train the adversarial intent recognition model. Construct a deep neural network that includes a feature embedding layer, a multi-path parallel feature extraction layer, and a feature fusion layer. At the same time, use the mined association rule results [state, action, time → intent] to filter adversarial intent. Step 4: Output the results and probabilities of the adversarial intent, and at the same time, use association rules to trace the source of the results to obtain the path of adversarial intent identification, so as to provide a basis for commanders' decision-making and assist in decision-making.

2. The spacecraft adversarial intent recognition method with interpretability according to claim 1, characterized in that, The temporal state data mentioned in step 1 is represented as: a set of four-tuples: [state, action, intention, next moment's state] in, It represents spacecraft status data, including attitude information, orbital information, and fuel information; Indicates an action command. For specific intent, The state at the next moment; Let the state vector of the principal spacecraft (i.e., the object of the identification intent) numbered B at time t be: in, This represents the x-coordinate and y-coordinate of spacecraft B on a two-dimensional plane. This indicates its lateral and longitudinal velocities on the two-dimensional orbital plane. For remaining fuel resources, This is the load condition.

3. The spacecraft adversarial intent recognition method with interpretability according to claim 2, characterized in that, In step 1, the preprocessing specifically includes: Step 1-1: Record the state of each spacecraft at time t at a fixed frequency from the simulation platform or remote sensing observation system. Then, preprocess the sample set and calculate the mean of each time series state data. and standard deviation Eliminate those that meet the requirements For outlier samples, linear interpolation is used to fill in the missing values ​​for a small number of missing measurements: Step 1-2: To address the class imbalance problem in the dataset, oversample the minority class.

4. The spacecraft adversarial intent recognition method with interpretability according to claim 2, characterized in that, Step 2 specifically involves processing the historical data, discretizing the time-series state data D to obtain discrete data Z, and then using the Apriori algorithm to mine frequent itemsets. And generate a set of association rules: in, , For minimum support, This represents the minimum confidence level.

5. The spacecraft adversarial intent recognition method with interpretability according to claim 3, characterized in that, The feature embedding layer described in step 3 uses a one-dimensional convolutional neural network to extract local spatial features.

6. The spacecraft adversarial intent recognition method with interpretability according to claim 5, characterized in that, The extraction of local spatial features specifically involves: the input of a one-dimensional convolutional neural network is preprocessed temporal state data, and multiple convolutional kernels of different widths are used in parallel to capture the local dependencies and patterns of features at different time scales; at the same time, during the model training process, the results of association rule mining [state, action, time → intent] are introduced in the form of constraints or auxiliary loss terms to perform logical filtering and consistency correction on the intent prediction output by the model.

7. The spacecraft adversarial intent recognition method with interpretability according to claim 4, characterized in that, The multi-path parallel feature extraction layer is divided into parallel branch A and parallel branch B. Parallel branch A adopts an improved bidirectional gated recurrent unit and introduces a temporal attention mechanism after the BiGRU layer to focus on the features of key time steps and capture local temporal dynamic changes. Parallel branch B adopts an improved Transformer encoder, which uses a multi-head self-attention mechanism for parallel computation. It retains the original encoder part of the Transformer and replaces the original decoder part with a prediction head. The Transformer can capture long-distance temporal dependencies and global contextual information.

8. The spacecraft adversarial intent recognition method with interpretability according to claim 5, characterized in that, In step 4, the final spacecraft adversarial intent and probability are output through a fully connected layer and a Softmax classifier. At the same time, the set of association rules mined in step 2 is used to reverse trace the predicted adversarial intent to generate a traceable reasoning path. The path can provide auxiliary decision-making information for the commander's situation assessment and decision-making.

9. A spacecraft adversarial intent recognition method with interpretability according to claim 6, characterized in that, The intent categories include reconnaissance, surveillance, guidance, interception, and escape.

10. A spacecraft adversarial intent recognition method with interpretability according to claim 1, characterized in that, This method is applicable to constrained game spaces.

Images