Knowledge graph-based target comprehensive intention reasoning method and device

Abstract

The application discloses a kind of target comprehensive intention reasoning method and device based on knowledge graph.The method comprises the following steps: constructing the knowledge graph of spacecraft;Training the graph neural network and long short-term memory network constructed in advance based on the knowledge graph, to obtain the trained graph neural network and long short-term memory network;Obtain the historical behavior data of target spacecraft and our spacecraft at multiple time steps;Match the historical behavior data with the knowledge graph, to obtain the corresponding spacecraft entity;With each spacecraft entity as node, the relationship between each entity as edge, based on the graph neural network, the characteristics of each node are aggregated and updated, to obtain the high-dimensional embedding vector of each node at each time step;Organize the historical behavior data into time series, and input it and each high-dimensional embedding vector into long short-term memory network, to obtain the comprehensive intention of target spacecraft;Based on the pre-constructed interactive interface, the comprehensive intention is output.The application can accurately infer the comprehensive intention of target.

Description

Technical Field

[0001] This invention relates to the field of intent recognition technology, and in particular to a method and apparatus for comprehensive intent reasoning based on knowledge graphs. Background Technology

[0002] With the rapid development of space technology and the increasing frequency of international space activities, the number of non-cooperative targets near Earth's low Earth orbit has grown dramatically over time, leading to a surge in the risk of rendezvous and collisions between spacecraft and other non-cooperative spacecraft with active maneuvering capabilities. Due to the lack of cooperation, accurately tracking and predicting the orbits of non-cooperative targets is difficult, and the inability to quickly and effectively respond to anomalies could result in severe losses. Therefore, to ensure the safe and stable operation of spacecraft, it is necessary to conduct advance reasoning studies on the comprehensive intentions of targets.

[0003] Currently, existing methods mainly assess the target's main intentions by measuring the target's motion relative to our spacecraft. However, this method has many problems, such as multiple measurement data modes, small sample size, numerous threat factors, complex correlations, and varied intentions, making it unable to quickly and accurately determine the target's comprehensive intentions. Summary of the Invention

[0004] This invention provides a method and apparatus for comprehensive intent reasoning based on knowledge graphs, which can accurately infer the comprehensive intent of a target. The technical solution is as follows: On the one hand, a target comprehensive intent reasoning method based on knowledge graph is provided, the method comprising: Construct a knowledge graph for spacecraft; the entities in the knowledge graph include the orbital information and motion status of various spacecraft, as well as the behavior and intentions of the various components within the spacecraft. Based on the knowledge graph, a pre-constructed graph neural network and a long short-term memory network are trained to obtain a trained graph neural network and a long short-term memory network. Acquire historical behavior data of the target spacecraft and our spacecraft at multiple time steps; The historical behavior data is matched with the knowledge graph to obtain the corresponding spacecraft entities; Using each spacecraft entity as a node and the relationship between each entity as an edge, the graph neural network is used to aggregate and update the features of each node to obtain a high-dimensional embedding vector for each node at each time step. The historical behavior data is organized into a time series, and then input into the long short-term memory network along with each of the high-dimensional embedding vectors to obtain the comprehensive intent of the target spacecraft. The comprehensive intent is output based on a pre-built interactive interface.

[0005] On the other hand, a target comprehensive intent reasoning device based on a knowledge graph is provided, the device comprising: A construction unit is used to build a knowledge graph of spacecraft; the entities in the knowledge graph include the orbital information and motion status of various spacecraft, as well as the behavior and intentions of various components in the spacecraft. The training unit is used to train the pre-constructed graph neural network and long short-term memory network based on the knowledge graph to obtain the trained graph neural network and long short-term memory network. The acquisition unit is used to acquire historical behavioral data of the target spacecraft and our spacecraft at multiple time steps. A matching unit is used to match the historical behavior data with the knowledge graph to obtain the corresponding spacecraft entity; The high-dimensional embedding vector determination unit is used to aggregate and update the features of each node based on the graph neural network, with each spacecraft entity as a node and the relationship between each entity as an edge, to obtain the high-dimensional embedding vector of each node at each time step. The comprehensive intent determination unit is used to organize the historical behavior data into a time series and input it and each of the high-dimensional embedding vectors into the long short-term memory network to obtain the comprehensive intent of the target spacecraft. The output unit is used to output the comprehensive intent based on a pre-built interactive interface.

[0006] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing computer programs, and the processor for executing the computer programs stored in the memory to implement the steps of the knowledge graph-based target comprehensive intent reasoning method described above.

[0007] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the above-described knowledge graph-based target comprehensive intent reasoning method.

[0008] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the knowledge graph-based target comprehensive intent reasoning method described above.

[0009] This invention provides a method for comprehensive intent reasoning based on knowledge graphs. By constructing a knowledge graph, historical information from different data sources, such as satellite on-orbit data, historical behavior records, and environmental factors, can be integrated to form a comprehensive knowledge system, enhancing the model's understanding of spacecraft behavior. By constructing a graph neural network, complex relationships between entities can be extracted, and high-dimensional embedding vectors can be output based on these extracted relationships, effectively representing the static and dynamic relationships between nodes. Finally, using the high-dimensional embedding vectors and the spacecraft's historical behavior data sequence as input to a long short-term memory network can improve the model's understanding and prediction capabilities of spacecraft intent, resulting in a more accurate comprehensive intent output. Therefore, this application provides reliable decision support for spacecraft behavior intent reasoning tasks, contributing to improved safety and efficiency of space missions and promoting the development and application of intelligent aerospace technology. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a target comprehensive intent reasoning method based on knowledge graphs provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a target comprehensive intent reasoning device based on a knowledge graph, provided in an embodiment of the present invention. Figure 3 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention; Figure 4 This is a spacecraft knowledge graph construction process provided in one embodiment of the present invention; Figure 5 This is a partial visualization diagram of a spacecraft knowledge graph provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a spacecraft structure extracted using the method described in this application, provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of parameters included in the historical behavior data of a spacecraft provided in an embodiment of the present invention; Figure 8 This is a flowchart of question intent recognition provided in an embodiment of the present invention; Figure 9 This refers to the loss variation during the training process of the question intent recognition model provided in an embodiment of the present invention. Figure 10This refers to the recall rate change during the training process of the question intent recognition model provided in an embodiment of the present invention. Figure 11 This refers to the change in accuracy during the training process of the question intent recognition model provided in an embodiment of the present invention. Figure 12 This is the F1 score change during the training process of the question intent recognition model provided in an embodiment of the present invention; Figure 13 This is an embodiment of the present invention providing the interaction flow between user requests and a graph knowledge reasoning system. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0013] The specific implementation of the method in this application is described in detail below.

[0014] Please refer to Figure 1 This invention provides a target comprehensive intent reasoning method based on knowledge graphs, the method comprising: Step 100: Construct a knowledge graph of the spacecraft; the entities in the knowledge graph include the orbital information, motion status, and behavior and intentions of various components of the spacecraft. Step 102: Train the pre-constructed graph neural network and long short-term memory network based on the knowledge graph to obtain the trained graph neural network and long short-term memory network; Step 104: Obtain historical behavior data of the target spacecraft and our spacecraft at multiple time steps; Step 106: Match historical behavior data with the knowledge graph to obtain the corresponding spacecraft entities; Step 108: Using each spacecraft entity as a node and the relationship between each entity as an edge, perform feature aggregation and update on each node based on a graph neural network to obtain a high-dimensional embedding vector for each node at each time step. Step 110: Organize the historical behavior data into a time series, and input it and each high-dimensional embedding vector into a long short-term memory network to obtain the comprehensive intent of the target spacecraft; Step 112: Output the comprehensive intent based on the pre-built interactive interface.

[0015] In this embodiment, by constructing a knowledge graph, historical information from different data sources, such as satellite on-orbit data, historical behavior records, and environmental factors, can be integrated to form a comprehensive knowledge system, enhancing the model's understanding of spacecraft behavior. By constructing a graph neural network (GNN), complex relationships between entities can be extracted, and high-dimensional embedding vectors can be output based on these extracted relationships, effectively representing the static and dynamic relationships between nodes. Finally, using the high-dimensional embedding vectors and the spacecraft's historical behavior data sequence as input to a long short-term memory network (LSTM) can improve the model's understanding and prediction capabilities of the spacecraft's intentions, outputting a more accurate comprehensive intention. Therefore, this application provides reliable decision support for spacecraft behavior intention reasoning tasks, contributing to improved safety and efficiency of space missions and promoting the development and application of intelligent aerospace technology.

[0016] The following description Figure 1 The execution method of each step is shown.

[0017] First, for step 100, construct a knowledge graph of the spacecraft, including: A1, crawling multi-source raw data of spacecraft; A2 preprocesses the raw data and performs word segmentation, part-of-speech tagging, and named entity recognition on the preprocessed data to obtain data text that meets the requirements. A3, extract target keywords related to intent recognition from the spacecraft keyword dictionary; A4, concatenate the target keywords with the data text, and use the set tags to annotate the target keywords to obtain the training dataset; A5, based on the training dataset, train the BERT model to obtain the trained BERT model; A6 inputs the collected historical behavior data of various spacecraft and target keywords into the trained BERT model, so as to extract entities related to spacecraft and the relationships between entities based on the BERT model, and summarize and organize the extracted entities and relationships to obtain the knowledge graph of spacecraft.

[0018] like Figure 4 The diagram illustrates the knowledge graph construction process. The following sections provide a detailed explanation of each step: For step A1, the multi-source raw data is obtained by sending requests to data websites to crawl historical and in-orbit spacecraft descriptive text snippets; the text snippets include basic spacecraft information, orbital parameters, component information, and behavioral intent characteristics, among which: Basic information about a spacecraft includes at least its name, category, nationality, and mission. The orbital parameters must include at least the orbital type, orbital inclination, and orbital period; Component information includes at least the solar panels, engine, and antenna; Behavioral and intentional characteristics include at least approaching, moving away, monitoring, and remaining still.

[0019] It should be noted that the types of components, behaviors, and intentions can include even more, as shown in Table 1: Table 1 Classification of Spacecraft Latent Intent Reasoning It should be noted that this application is not limited to the contents shown in Table 1.

[0020] The following details the process of crawling raw data from multiple sources: This project uses Python to write a web crawler program. To build a multi-threaded spacecraft information crawler, the `requests` and `BeautifulSoup` libraries are used. A crawler class, `SpacecraftCrawler`, is created, which initializes a list of URLs to crawl and a list to store the results. In the `fetch_url` method, `requests.get` is used to send HTTP requests, setting a timeout of N seconds (e.g., 10 seconds), and the `raise_for_status()` method is used to check if the request was successful. If the request is successful, the webpage content is returned. Next, the `parse_content` method is defined, using BeautifulSoup to parse the HTML document and extract basic information about the spacecraft, including orbital parameters, relative trajectory type, components, behaviors, and intents. In the main crawler logic `crawl`, a `ThreadPoolExecutor` is used, with a maximum number of threads (e.g., 5) to process multiple requests in parallel. The `executor.submit()` method submits the requests to the thread pool, and `future.result()` is used to retrieve the crawling results. The successfully obtained HTML content is passed to parse_content for parsing, and the parsed information is stored in the results list. Finally, the crawled spacecraft information is output and stored in a CSV file. An error handling mechanism is added to ensure that network errors or parsing errors can be captured during the request and parsing process. Request frequency control is also considered to prevent excessive pressure on the target website, thereby reducing the risk of being blocked.

[0021] For step A2, preprocessing the raw data includes cleaning and formatting the raw data, and removing HTML tags, redundant spaces and special characters using regular expressions.

[0022] After preprocessing, the spaCy natural language processing library is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the original text segments. Furthermore, to ensure the diversity of training samples for the entity extraction model, data augmentation methods such as synonym replacement, random insertion, and random deletion are employed after information collection to increase data diversity.

[0023] For step A3, the target keywords were determined as follows: Determine the total number of documents in the spacecraft keyword dictionary; For each word in the spacecraft keyword dictionary, calculate the inverse document frequency of that word in the entire keyword dictionary; For each document, calculate the frequency of each word in the document; and use the product of the word's frequency and its inverse document frequency as the word's overall weight; sort the overall weights in descending order, and select a set number of words with larger overall weights as the target keywords for the document.

[0024] In this step, the inverse document frequency of a word is calculated using the following formula: The frequency of words in a document is calculated using the following formula: The overall weight of words is calculated using the following formula: In the formula, Words t In the document d The frequency of occurrence in words, i.e., word frequency; For words t The entire collection of information documents D Inverse document frequency; Words t In the document d The overall weighting in; N Total number of documents; Indicates any word; Words In the document d Frequency of occurrence in.

[0025] For step A4, the target keyword is concatenated with the data text, and a [KEYWORD] marker is used to enhance the model's focus on the target keyword. In addition to the traditional input ID and token type ID, extra embedding is added to the target keyword and combined with the data text embedding.

[0026] This step, by adding keywords, enhances the input features, enabling the BERT model to better capture key contextual information points when processing spacecraft-related linguistic information, thereby summarizing entities and relationships in the spacecraft dataset.

[0027] In addition, target keywords include: spacecraft type, mission, components, behavior, intent, orbit type, relative trajectory type, etc. Of course, users can also set other keywords as needed, and this application does not impose specific limitations.

[0028] In addition, the spacecraft dataset contains several key entity attributes, which will be systematically organized to facilitate subsequent analysis and inference. These entity attributes include: a) Basic Information – Spacecraft Overview; b) Time – Spacecraft in orbit; c) Nationality – The country to which the spacecraft belongs, used to determine whether the action is cooperative or confrontational; d) Category – Spacecraft category, serving as the basis for comprehensive analysis of its intent; e) Mission – The primary operational purpose of the spacecraft; f) Orbital information—orbital attributes for stable spacecraft operation, including orbital type, perigee altitude, apogee altitude, orbital inclination, orbital period, etc. g) Trajectory type – the relative trajectory of the target spacecraft and a specific spacecraft; h) Components – Major components on a spacecraft; i) Behavior – Typical behavior of spacecraft components; j) Intent – ​​The intent behind the typical behavior of a spacecraft component; k) Comprehensive Intent – ​​A comprehensive intent derived from the combined inferences of the intents of spacecraft components, including independent intent, cooperative intent, and adversarial intent. Independent integration means that the spacecraft can perform its mission autonomously without interacting with other spacecraft. The overall intent of the collaboration is for spacecraft to operate in tandem with other spacecraft; The overall intent to counter the threat posed by the spacecraft is to engage in threatening behavior.

[0029] For step A5, regularly monitor the metrics related to the keywords (such as keyword recognition accuracy, recall rate, etc.). If the recognition effect of similar entities of a certain keyword is not good, increase the sample size of the keyword or adjust the attention weight, and set an early stopping mechanism. If the keyword recognition performance on the validation set does not improve within several consecutive epochs, stop training to prevent overfitting and ensure that the model has strong generalization ability in entity recognition.

[0030] For step A6, after the BERT model is trained, it can be used to generate a knowledge graph. At this point, it's necessary to collect a larger-scale dataset of historical behavior data from various spacecraft. The collected information and target keywords are then input into the BERT model to obtain the knowledge graph, such as... Figure 5 The image shown is a partial visualization of the spacecraft knowledge graph.

[0031] Furthermore, the spacecraft knowledge graph constructed in this application logically comprises a data layer and a schema layer. The data layer consists of a series of facts; knowledge is stored in units of facts, and facts are stored as triples of "entity-relationship-entity" or "entity attribute-attribute value." Entities are the basic elements of the knowledge graph; specifically in the spacecraft knowledge graph, this refers to each specific spacecraft name, mission, orbit type, optical payload, etc. Relationships are the semantic relationships between two entities, and attribute values ​​are descriptions of the entities. The schema layer, built upon the data layer, primarily focuses on the structure and semantic definition of the knowledge graph. The conceptual structure of entities, relationships, and attributes is defined at this layer, providing an abstract way to describe the logical structure of the knowledge graph.

[0032] It should also be noted that after the knowledge graph is built, it can be connected to the local Neo4j database and authenticated. Functions for node management and relationship creation are defined. In the main program, the database is first cleared to ensure a clean environment. Then, the spacecraft entity relationship dictionary is loaded from a text file, the dataset is read to extract spacecraft-related information, entity information is extracted line by line, and nodes and relationships are created using custom functions to realize the visual construction of the comprehensive spacecraft knowledge graph.

[0033] For step 102, the following are included: A sample set consisting of historical on-orbit data is constructed. Each sample in the sample set is labeled with a high-dimensional embedding vector label corresponding to each entity node and a comprehensive intent label of the sample. Each sample is input into the trained BERT model to identify keywords and entities in the sample. The identification results are then matched with the knowledge graph to obtain the corresponding entities and the relationships between them. The obtained entities are used as nodes, and the relationships between entities are used as edges. Each node and edge is used as the input to the graph neural network. The graph neural network is trained based on the high-dimensional embedding vector labels of each node until the model converges, resulting in a trained graph neural network. Each sample is organized into a time series, and each time series and the output of the trained graph neural network are used as inputs to the long short-term memory network. The long short-term memory network is trained based on the comprehensive intent label of the samples until the model converges, resulting in a trained long short-term memory network.

[0034] By training graph neural networks and long short-term memory networks, the graph neural network can aggregate features of nodes, extract information from neighboring nodes, generate updated node embedding representations, and output a high-dimensional embedding vector for each node. This enables the long short-term memory network to output the spacecraft's latent intent within a given time step.

[0035] In some implementations, the graph neural network is updated during training using the following formula: In the formula, For nodes v In the k The feature vector of the layer; N( is the activation function;) v ) is a node v The neighboring nodes; For the first k Layer weight matrix; For nodes u and nodes v The weight of the edges between them; For nodes v The normalization factor; For nodes u In the k Feature vectors of layer -1.

[0036] Through the above update method, graph neural networks can extract information from neighboring nodes and generate updated node embedding representations.

[0037] Furthermore, during the training of the graph neural network, the Adam optimizer was used with a learning rate of 0.001 and a batch size of 64, and training was performed for 100 epochs. The cross-entropy loss function was chosen, with the following formula: in, For real labels, For the model's predicted values, Dropout regularization is used to prevent overfitting, with the Dropout ratio set to 0.5.

[0038] In some implementations, the Long Short-Term Memory (LSTM) network includes cell states and hidden states, with the hidden states updated using the following formula. h t and unit state c t : In the formula, i t , f t , o t They are respectively t Input gate, forget gate, and output gate at time steps. , , These represent the weights of the input gate, forget gate, output gate, and cell state, respectively. , , , These represent the biases of the input gate, forget gate, output gate, and cell state, respectively. The candidate cell state is calculated using the tanh activation function and is used to generate the cell states that may be added at the current time step. c t New information in the middle. It will be combined with the input gate i t The output of the algorithm determines how much new information to add to the unit state. Through this update method, the Long Short-Term Memory (LSTM) network can capture potential dependencies in time series data. During LSTM network training, the Adam optimizer was used with a learning rate of 0.001 and a batch size of 128, for 200 epochs. An attention mechanism was also introduced to improve the performance of the LSTM on time series data.

[0039] In some implementations, the probability distribution of the overall intent output by the Long Short-Term Memory network can also be calculated using the following formula: In the formula, It is the output weight matrix. These are the node feature vectors generated by the GNN model (representing the spacecraft's feature information in the graph structure). It is a time-series feature vector generated by the LSTM model, which captures the behavioral changes of the spacecraft in the time dimension; It is a combination of information such as "component-behavior-intention" in an event. i It is a positive integer belonging to [1, m], where m is the total number of information combination events; The "comprehensive intention in the analysis of the event is the actual intention" jIt is a positive integer belonging to [1, n], where n is the total number of potential integrated intentions. By fusing across models, the ability and accuracy of spacecraft behavior recognition are enhanced, and the integrated intention reasoning of spacecraft behavior will be more accurate and efficient, better supporting integrated interaction and decision-making. In this way, LSTM can capture potential dependencies in the time series, and the output is the probability distribution of the predicted behavioral intention, representing the potential intention of the spacecraft in a given time step. The specific classification of potential intentions is shown in Table 1. After completing the temporal reasoning, the path of the reasoning matching result in the knowledge graph is saved for subsequent interactive calls.

[0040] Regarding step 104, such as Figure 7 As shown, historical behavior data can include long-range orbital information, medium-range trajectory, short-range motion and component status, target type, identification intent, and target attributes. Among these, the most important are the spacecraft's components, component status, performance, trajectory, instantaneous position, velocity, and acceleration.

[0041] For step 106, some non-standardized descriptive words in the historical behavior data can be matched into standardized words to meet the input requirements of the graph neural network.

[0042] For step 108, the historical behavior data includes multiple time steps. Taking data within one minute as an example, if each time stamp is 5 seconds, there are a total of 12 time steps. Each time step outputs a high-dimensional embedding vector to represent the static and dynamic relationships between nodes. Static relationships include orbital altitude difference, orbital inclination difference, etc.; dynamic relationships include the spacecraft's behavior in the next time step, such as orbit ascent or docking. The output of the graph neural network is the spacecraft's short-term intent, that is, the intent of a specific behavior of a certain component.

[0043] For step 110, the input to the Long Short-Term Memory network includes a high-dimensional embedding vector output by the graph neural network and a time-series feature vector composed of historical behavioral data, including multiple parameters at three distances (far, medium, and near) of the spacecraft. By using the above composite input, more features can be learned, and the overall intent of the output is more accurate. Figure 6 As shown, the overall intent includes independent integration intent, cooperative integration intent, and confrontational integration intent. Among them: The independent integration intent is for the target spacecraft to autonomously execute its mission without interacting with other spacecraft; The overall intent of the collaboration is for the target spacecraft to operate in coordination with other spacecraft. The overall intent of the countermeasure is to counter the threatening behavior of the target spacecraft against other spacecraft.

[0044] It should be noted that other spacecraft can be our own spacecraft, or spacecraft other than the target spacecraft that our spacecraft observes.

[0045] Finally, regarding step 112, to enable interaction with the knowledge graph and inference results, a backend service is built using the Django framework to interact with the knowledge graph via an HTTP interface. URL routing and view functions are configured in the Django application, and a RESTful API is designed to support the reception and processing of user query requests. Relevant information is extracted from the knowledge and inference result paths through a Neo4j database connection, and a visual interface enables efficient interaction between user requests and the knowledge graph inference system.

[0046] The following specific embodiments illustrate the interaction process that defines the overall intent of this application: Assuming the received historical behavioral data of both the target spacecraft and our spacecraft within a set time period are as follows: The orbits of GSSAP-1 and Kosmos-2560 have a small inclination difference. GSSAP-1's orbit relative to Kosmos-2560 is a medium-amplitude high-orbit orbit. GSSAP-1's trajectory relative to Kosmos-2560 is a teardrop-shaped forward spiral. GSSAP-1's solar panels close to assist in orbital maneuvers. GSSAP-1's satellite engines fire multiple times in multiple directions to adjust its attitude. GSSAP-1's satellite engines fire in opposite directions for extended periods to perform a large-scale orbital descent. GSSAP-1's radar remote sensing instrument rotates towards the star to measure the distance between them.

[0047] Based on the behavioral data above, the question is: What are the components, behaviors, intentions, and overall intentions of GSSAP-1? The aforementioned behavioral data is then matched with the knowledge graph. Based on the matching results, the GNN outputs high-dimensional embedding vectors for each node, such as orbital inclination difference: small inclination difference; orbital altitude difference: medium amplitude high orbit; relative trajectory type: water droplet forward spiral; solar panel - closure - auxiliary orbit change: engine - reverse long-term ignition - large orbit reduction; inter-satellite ranging; engine - multi-directional ignition - attitude adjustment.

[0048] By inputting the above high-dimensional embedding vectors and behavioral data into the LSTM, the following comprehensive intent matching result can be obtained: small tilt difference + medium amplitude high orbit + water droplet forward hovering + auxiliary orbit change + attitude adjustment + large orbit reduction + inter-satellite ranging = high orbit difference same-direction approach hovering.

[0049] It should also be noted that this application can also assess user intent. When user intent assessment is required, a BERT-Base-Chinese model, i.e., a question intent recognition model, needs to be constructed first. A training dataset is constructed, and various question styles are designed to cover different inquiry intents. The training samples are shown in Table 2, which provides multiple question classification examples to ensure data diversity.

[0050] Table 2 Examples of Classification of Spacecraft Domain Knowledge and Behavioral Intent Reasoning Problems Then, a masked language model (MLM) is used for training, combined with a Transformer encoder to extract contextual information of sentences, in order to enhance the model's ability to understand natural language.

[0051] The dataset partitioning steps include: before inputting the training data into the model, all data is randomly shuffled to avoid bias caused by order. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio, containing 2909, 363, and 365 data points respectively. This partitioning method ensures the effectiveness of the training process and the generalization ability of the model, which helps to improve the model's performance in real-world application scenarios. The BERT model is fine-tuned, and the optimal hyperparameter settings during the model training phase are shown in Table 3, including maximum sentence length (256), batch size (16), learning rate (5e-5), loss rate (0.1), and number of training epochs (20).

[0052] Table 3 Optimal training parameters for the question intent recognition model During training, a Dropout layer is used to randomly deactivate some neurons to prevent overfitting and improve the model's generalization ability. Finally, a Softmax layer is used for classification, normalizing the score for each category and outputting the probability of each category to effectively distinguish between different question intents. A flowchart illustrating the question intent recognition process is shown below. Figure 8 As shown in the figure. During model training, multiple metrics were used to evaluate model performance, and a residual monitor was set up. The changes in loss, recall, precision, and F1 score during training are shown in the figure. Figure 9 , Figure 10 , Figure 11 , Figure 12 As shown.

[0053] By adopting the above model, such as Figure 13 As shown, when a user poses an intent question, the question is input into the aforementioned question intent recognition model, which can identify the entities and intents in the question. The recognition results are then matched with the knowledge graph to obtain the corresponding answer, which is then output to the front-end page.

[0054] like Figure 2 , Figure 3 As shown, this embodiment of the invention provides a target comprehensive intent reasoning device based on a knowledge graph. The device embodiment can be implemented through software, hardware, or a combination of both. From a hardware perspective, such as... Figure 2The diagram shown is a hardware architecture diagram of a computing device housing a knowledge graph-based target comprehensive intent reasoning device according to an embodiment of the present invention. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, the computing device in the embodiment may also include other hardware, such as a forwarding chip responsible for processing packets. Taking software implementation as an example, such as... Figure 3 As shown, a device in a logical sense is formed by the CPU of the computing device in which it is located reading the corresponding computer program from the non-volatile memory into the memory for execution.

[0055] Please refer to Figure 3 This invention provides a target comprehensive intent reasoning device based on knowledge graphs, the device comprising: Building unit 300 is used to build a knowledge graph of spacecraft; the entities in the knowledge graph include the orbital information and motion status of various spacecraft, as well as the behavior and intentions of various components in the spacecraft. Training unit 302 is used to train a pre-constructed graph neural network and long short-term memory network based on a knowledge graph, so as to obtain a trained graph neural network and long short-term memory network. Acquisition unit 304 is used to acquire historical behavior data of the target spacecraft and our spacecraft at multiple time steps; Matching unit 306 is used to match historical behavior data with the knowledge graph to obtain the corresponding spacecraft entity; The high-dimensional embedding vector determination unit 308 is used to aggregate and update the features of each node based on a graph neural network, with each spacecraft entity as a node and the relationship between each entity as an edge, to obtain the high-dimensional embedding vector of each node at each time step. The comprehensive intent determination unit 310 is used to organize historical behavior data into a time series and input it and each high-dimensional embedding vector into a long short-term memory network to obtain the comprehensive intent of the target spacecraft. Output unit 312 is used to output a comprehensive intent based on a pre-built interactive interface.

[0056] In some implementations, the construction unit 300 is used to perform the following operations: Crawling multi-source raw data from spacecraft; The raw data is preprocessed, and then the preprocessed data is segmented, part-of-speech tagging is performed, and named entity recognition is performed to obtain data text that meets the requirements; Extract target keywords related to intent recognition from the spacecraft keyword dictionary; The target keywords are concatenated with the data text, and the target keywords are labeled using defined tags to obtain the training dataset; The BERT model is trained based on the training dataset to obtain a trained BERT model; The collected historical behavior data of various spacecraft and target keywords are input into the trained BERT model to extract entities related to spacecraft and the relationships between these entities. The extracted entities and relationships are then summarized and organized to obtain a knowledge graph of the spacecraft.

[0057] In some implementations, the training unit 302 is used to perform the following operations: A sample set consisting of historical on-orbit data is constructed. Each sample in the sample set is labeled with a high-dimensional embedding vector label corresponding to each entity node and a comprehensive intent label of the sample. Each sample is input into the trained BERT model to identify keywords and entities in the sample. The identification results are then matched with the knowledge graph to obtain the corresponding entities and the relationships between them. The obtained entities are used as nodes, and the relationships between entities are used as edges. Each node and edge is used as the input to the graph neural network. The graph neural network is trained based on the high-dimensional embedding vector labels of each node until the model converges, resulting in a trained graph neural network. Each sample is organized into a time series, and each time series and the output of the trained graph neural network are used as inputs to the long short-term memory network. The long short-term memory network is trained based on the comprehensive intent label of the samples until the model converges, resulting in a trained long short-term memory network.

[0058] In some implementations, the multi-source raw data is obtained by sending requests to data websites to crawl historical and in-orbit descriptive segments of spacecraft; the descriptive segment information includes basic information about the spacecraft, orbital parameters, component information, and behavioral intent characteristics, wherein: Basic information about a spacecraft includes at least its name, category, nationality, and mission. The orbital parameters must include at least the orbital type, orbital inclination, and orbital period; Component information includes at least the solar panels, engine, and antenna; Behavioral and intentional characteristics include at least approaching, moving away, monitoring, and remaining still.

[0059] In some implementations, the target keywords are determined in the following way: Determine the total number of documents in the spacecraft keyword dictionary; For each word in the spacecraft keyword dictionary, calculate the inverse document frequency of that word in the entire keyword dictionary; For each document, calculate the frequency of each word in the document; and use the product of the word's frequency and its inverse document frequency as the word's overall weight; sort the overall weights in descending order, and select a set number of words with larger overall weights as the target keywords for the document.

[0060] In some implementations, target keywords include: spacecraft type, mission, component, behavior, intent, orbit type, and relative trajectory type.

[0061] In some implementations, the integration intent includes independent integration intent, cooperative integration intent, and adversarial integration intent; The independent integration intent is for the target spacecraft to autonomously execute its mission without interacting with other spacecraft; The overall intent of the collaboration is for the target spacecraft to operate in coordination with other spacecraft. The overall intent of the countermeasure is to counter the threatening behavior of the target spacecraft against other spacecraft.

[0062] It should be noted that the knowledge graph-based target intent reasoning device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the knowledge graph-based target intent reasoning device and the knowledge graph-based target intent reasoning method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0063] Embodiments of this application also provide a computer device, please refer to... Figure 3 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set, or instruction set, the at least one instruction, at least one program, code set, or instruction set being loaded and executed by the processor to implement the knowledge graph-based target comprehensive intent reasoning method provided in the above-described method embodiments.

[0064] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the knowledge graph-based target comprehensive intent reasoning method provided in the above-described method embodiments.

[0065] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the knowledge graph-based target comprehensive intent reasoning methods described in the above embodiments.

[0066] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0067] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, 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 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 various embodiments or some parts of the embodiments of this application.

[0068] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0069] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A target comprehensive intent reasoning method based on knowledge graph, characterized in that, The method includes: Construct a knowledge graph for spacecraft; the entities in the knowledge graph include the orbital information and motion status of various spacecraft, as well as the behavior and intentions of the various components within the spacecraft. Based on the knowledge graph, a pre-constructed graph neural network and a long short-term memory network are trained to obtain a trained graph neural network and a long short-term memory network. Acquire historical behavior data of the target spacecraft and our spacecraft at multiple time steps; The historical behavior data is matched with the knowledge graph to obtain the corresponding spacecraft entities; Using each spacecraft entity as a node and the relationship between each entity as an edge, the graph neural network is used to aggregate and update the features of each node to obtain a high-dimensional embedding vector for each node at each time step. The historical behavior data is organized into a time series, and then input into the long short-term memory network along with each of the high-dimensional embedding vectors to obtain the comprehensive intent of the target spacecraft. The comprehensive intent is output based on a pre-built interactive interface; The step of training a pre-constructed graph neural network and a long short-term memory network based on the knowledge graph to obtain the trained graph neural network and long short-term memory network includes: Construct a sample set consisting of historical on-orbit data, wherein each sample in the sample set is labeled with a high-dimensional embedding vector label corresponding to each entity node and a comprehensive intent label of the sample; Each sample is input into the trained BERT model to identify keywords and entities in the sample. The identification results are then matched with the knowledge graph to obtain the corresponding entities and the relationships between them. The obtained entities are used as nodes, and the relationships between entities are used as edges. Each node and edge is used as the input to the graph neural network. The graph neural network is trained based on the high-dimensional embedding vector labels of each node until the model converges, resulting in a trained graph neural network. Each sample is organized into a time series, and each time series and the output of the trained graph neural network are used as the input of the long short-term memory network. The long short-term memory network is trained based on the comprehensive intent label of the sample until the model converges, and the trained long short-term memory network is obtained.

2. The method according to claim 1, characterized in that, The knowledge graph for constructing spacecraft includes: Crawling multi-source raw data from spacecraft; The original data is preprocessed, and the preprocessed data is then segmented, part-of-speech tagging is performed, and named entity recognition is performed to obtain data text that meets the requirements. Extract target keywords related to intent recognition from the spacecraft keyword dictionary; The target keywords are concatenated with the data text, and the target keywords are labeled using a set tag to obtain the training dataset; The BERT model is trained based on the training dataset to obtain a trained BERT model; The collected historical behavior data of various spacecraft and target keywords are input into the trained BERT model to extract entities related to spacecraft and the relationships between these entities. The extracted entities and relationships are then summarized and organized to obtain a knowledge graph of the spacecraft.

3. The method according to claim 2, characterized in that, The multi-source raw data is obtained by sending requests to data websites to crawl historical and in-orbit spacecraft description segments; the segment information includes basic information about the spacecraft, orbital parameters, component information, and behavioral intent characteristics.

4. The method according to claim 2, characterized in that, The target keywords were determined in the following way: Determine the total number of documents in the spacecraft keyword dictionary; For each word in the spacecraft keyword dictionary, calculate the inverse document frequency of that word in the entire keyword dictionary; For each document, calculate the frequency of each word in that document; and use the product of the word's frequency and its inverse document frequency as the overall weight of that word. Each comprehensive weight is sorted in descending order, and the words with the largest number of comprehensive weights are selected as the target keywords for the document.

5. The method according to claim 2, characterized in that, The target keywords include: spacecraft type, mission, components, behavior, intent, orbit type, and relative trajectory type.

6. The method according to claim 1, characterized in that, The integrated intent includes independent integrated intent, cooperative integrated intent, and confrontational integrated intent; The independent integration intent is for the target spacecraft to perform its mission autonomously without interacting with other spacecraft; The overall intent of the cooperation is for the target spacecraft to operate collaboratively with other spacecraft. The intent of the countermeasure is that the target spacecraft poses a threat to other spacecraft.

7. A target comprehensive intent reasoning device based on knowledge graph, characterized in that, The apparatus for implementing the steps of any one of claims 1-6, the apparatus comprising: A construction unit is used to build a knowledge graph of spacecraft; the entities in the knowledge graph include the orbital information and motion status of various spacecraft, as well as the behavior and intentions of various components in the spacecraft. The training unit is used to train the pre-constructed graph neural network and long short-term memory network based on the knowledge graph to obtain the trained graph neural network and long short-term memory network. The acquisition unit is used to acquire historical behavioral data of the target spacecraft and our spacecraft at multiple time steps. A matching unit is used to match the historical behavior data with the knowledge graph to obtain the corresponding spacecraft entity; The high-dimensional embedding vector determination unit is used to aggregate and update the features of each node based on the graph neural network, with each spacecraft entity as a node and the relationship between each entity as an edge, to obtain the high-dimensional embedding vector of each node at each time step. The comprehensive intent determination unit is used to organize the historical behavior data into a time series and input it and each of the high-dimensional embedding vectors into the long short-term memory network to obtain the comprehensive intent of the target spacecraft. The output unit is used to output the comprehensive intent based on a pre-built interactive interface.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

Images