A target spacecraft intention recognition method and system based on a large language model

Abstract

The application discloses a target spacecraft intention recognition method and system based on a large language model, and belongs to the technical field of aerospace. The method comprises the following steps: constructing a target spacecraft intention library; based on the target spacecraft intention library, constructing a prompt element for interaction with a large language model, and generating a prompt word corpus; training the large language model according to the prompt word corpus to obtain a trained large language model; and recognizing the target spacecraft intention by using the trained large language model. The application realizes automatic intention judgment and reasoning in the spacecraft movement process based on the causal reasoning capability of the large language model.

Description

Technical Field

[0001] This invention relates to the field of aerospace technology, and more specifically to a method and system for target spacecraft intent recognition based on a large language model. Background Technology

[0002] With the surge in the number of artificial satellites, the complexity of the space environment is indeed increasing, posing unprecedented challenges to space security. Space experiments by some countries have not only demonstrated their formidable technological capabilities but have also sparked discussions about the possibility of space warfare. Against this backdrop, early warning systems for spacecraft orbital safety are of paramount importance.

[0003] Traditional safety early warning methods primarily rely on distance assessments between satellites, but this approach has increasingly revealed its limitations. First, it neglects the relative motion configurations between spacecraft, which is crucial for accurately determining spacecraft trajectories and potential collision risks. Second, traditional methods often overlook critical information such as the type and purpose of the target spacecraft. Different types of spacecraft may have different flight trajectories and intentions, factors essential for accurately assessing potential threats.

[0004] Furthermore, with the increasing complexity of the space environment, both the quantity and quality of situational awareness data face enormous challenges. The difficulty of manually analyzing this data to infer target intentions is constantly increasing, especially given the explosive growth of data. Moreover, existing data often only provides one-sided information on relative orbital position and velocity, lacking comprehensiveness, which easily leads to misjudgments and missed detections.

[0005] Therefore, how to accurately determine the intentions of a target spacecraft by utilizing information from multiple sources is one of the important problems that urgently need to be solved in this field. Summary of the Invention

[0006] To address the problem in existing technologies of accurately determining the intentions of a target spacecraft using multiple sources of information, this invention provides a target spacecraft intention recognition method based on a large language model. Leveraging the causal reasoning capabilities of the large language model, it enables automatic intention judgment and reasoning during spacecraft motion.

[0007] To achieve the above objectives, the present invention provides the following technical solution.

[0008] A method for target spacecraft intent recognition based on a large language model includes:

[0009] Build a target spacecraft intent database;

[0010] Based on the target spacecraft intent database, construct prompting elements that interact with a large language model and generate a prompting word corpus;

[0011] Based on the prompt word corpus, the large language model is trained to obtain the trained large language model;

[0012] The trained large language model is used to identify the target spacecraft's intent.

[0013] As a further improvement of the present invention, the construction of the target spacecraft intent library is achieved by generating intent words under different intent attributes through a large language model, then analyzing and judging all the generated intent words, and organizing the correctly judged intent words to obtain the target spacecraft intent library.

[0014] As a further improvement of the present invention, the spacecraft intent library includes orbital intent, mission intent, and mission intent.

[0015] As a further improvement of the present invention, the step of constructing prompting elements that interact with a large language model based on the target spacecraft intent library and generating a prompting word corpus includes:

[0016] Based on the target spacecraft intent database, determine the information needed to obtain the corresponding intent in typical scenarios;

[0017] Based on the information needed to obtain the corresponding intent, construct prompt elements for interaction with the large language model;

[0018] The prompting elements for interaction with the large language model are expanded to obtain prompting vocabulary;

[0019] Organize the obtained prompt words and generate a prompt corpus.

[0020] As a further improvement of the present invention, the prompting elements for interaction with the large language model include the current situation, our information, target information, environmental conditions, and relative orbital motion characteristics.

[0021] As a further improvement of the present invention, the step of training the large language model based on the prompt word corpus to obtain the trained large language model includes:

[0022] Construct standard templates for prompt statements for input to a large language model based on a corpus of prompt words;

[0023] Sample data for input prompts are constructed based on prompt elements and standard templates for prompt statements in a large language model.

[0024] The large language model is trained by using sample data provided by the large language model as input, resulting in a trained large language model.

[0025] As a further improvement of the present invention, the step of constructing sample data for prompt input of the large language model based on prompt elements and prompt statement standard templates involves filling prompt words into prompt statement standard templates, and then filtering and evaluating possible intentions to obtain sample data for prompt input of the large language model.

[0026] A target spacecraft intent recognition system based on a large language model includes:

[0027] The intent library building module is used to build the intent library for the target spacecraft.

[0028] Corpus building module: used to build prompt elements that interact with the large language model based on the target spacecraft intent library, and generate a prompt word corpus;

[0029] Training Model Module: Used to train the large language model based on the prompt word corpus to obtain the trained large language model;

[0030] Intent recognition module: Used to identify the intent of the target spacecraft using a trained large language model.

[0031] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the target spacecraft intent recognition method based on a large language model.

[0032] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the target spacecraft intent recognition method based on a large language model.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This invention constructs a specialized target spacecraft intent database, ensuring that the large language model fully learns various possible spacecraft intents during training. This targeted learning enables the model to more accurately capture and interpret spacecraft intents in subsequent recognition processes. Furthermore, by designing cue elements that interact with the large language model and generating a cue word corpus, this invention allows for flexible adjustments to the model's input methods, making it more adaptable to the specific task of spacecraft intent recognition. This adaptability enables the model to handle spacecraft intent recognition problems in different scenarios and backgrounds. Secondly, the large language model used in this invention possesses powerful causal reasoning capabilities, capable of analyzing the correlations and causal relationships between various factors during spacecraft motion. Combined with judgments on orbital dynamics evolution, the model can gain a deeper understanding of the spacecraft's motion patterns and behavior, thereby more accurately determining its intent. Finally, through the trained large language model, automatic recognition of target spacecraft intents can be achieved. This automated and intelligent processing method significantly reduces the burden of manual analysis, improves processing efficiency and accuracy, and also reduces the risk of human error. Therefore, this invention, based on the causal reasoning capabilities of the large language model and combined with judgments on orbital dynamics evolution, achieves automatic intent judgment and reasoning during spacecraft motion. This not only improves the technical level of spacecraft intent recognition but also provides new ideas and methods for innovative development in the aerospace field. Through continuous optimization and improvement of the model, its applications in space mission planning, spacecraft control, and space situational awareness can be further expanded, driving the continuous progress of aerospace technology. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a target spacecraft intent recognition method based on a large language model according to the present invention.

[0036] Figure 2 This is a flowchart illustrating the overall steps of a target spacecraft intent recognition method based on a large language model according to the present invention.

[0037] Figure 3 This is a flowchart illustrating the specific steps of S1 in the target spacecraft intent recognition method based on a large language model of the present invention.

[0038] Figure 4 This is the spacecraft intent corpus constructed for the target spacecraft intent recognition method S1 based on a large language model of the present invention;

[0039] Figure 5 This is a flowchart illustrating the specific steps of S2 in the target spacecraft intent recognition method based on a large language model of the present invention.

[0040] Figure 6This is a corpus of prompting elements constructed for the target spacecraft intent recognition method S2 based on a large language model of the present invention. Detailed Implementation

[0041] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0043] To address the problem in existing technologies of accurately determining the intent of a target spacecraft using multiple pieces of information, this invention provides a target spacecraft intent recognition method based on a large language model, such as... Figure 1 As shown, the method is as follows:

[0044] Build a target spacecraft intent database;

[0045] Based on the target spacecraft intent database, construct prompting elements that interact with a large language model and generate a prompting word corpus;

[0046] Based on the prompt word corpus, the large language model is trained to obtain the trained large language model;

[0047] The trained large language model is used to identify the target spacecraft's intent.

[0048] This method leverages the causal reasoning capabilities of large language models to achieve automatic intent judgment and reasoning during spacecraft motion.

[0049] The present invention will be further explained and described below with reference to the accompanying drawings.

[0050] like Figure 2 As shown, a target spacecraft intent recognition method based on a large language model includes the following steps:

[0051] S1: As Figure 3 As shown, a target spacecraft intent library is constructed for typical scenarios;

[0052] S11: Different intent attributes in typical scenarios are divided into three categories: track intent, operation intent, and task intent.

[0053] S12: Generate intent vocabulary under different intent attributes based on different intent attributes in typical scenarios;

[0054] S13: It is also possible to generate intent vocabulary under different intent attributes through a large language model;

[0055] S14: To ensure the accuracy of intent words, all intent words generated in S12 and S13 need to be analyzed and judged multiple times, for example, by experts.

[0056] S15: Organize the intentional words correctly identified in S14, and obtain the following: Figure 4 The spacecraft shown is intended for use in a pre-planned storage facility;

[0057] In typical scenarios, spacecraft intent classification is achieved by analyzing different intent attributes.

[0058] The spacecraft intent library is divided into orbital intents, mission intents, and mission intents. Figure 3 kind.

[0059] Among these, orbital intent is classified based on relative orbital distance and relative velocity; operational intent is classified based on target size and onboard equipment; and mission intent is determined jointly by the spacecraft's orbital and operational intents.

[0060] The specific intent for each category is as follows:

[0061] (1) There are 7 types of intentions for orbital motion:

[0062] 1) Maintain: The relative orbital distance remains constant, and the relative velocity is zero;

[0063] 2) Skipping: The relative orbital distance changes continuously, but there is a minimum value;

[0064] 3) Accompanying or circling: The relative orbital distance varies within a relatively small range;

[0065] 4) Intersection: After orbital maneuvering, the relative orbital distance is zero, and the relative velocity is also zero;

[0066] 5) Deterrence: The difference from interception is that there are fewer target spacecraft, usually only one;

[0067] 6) Avoidance: Before the target orbital maneuver, the relative orbital distance will become very small in the future, which may lead to a collision. After the maneuver, the spacecraft will not collide.

[0068] 7) Interception: After the target maneuvers, the relative orbital distance will decrease in the future, leading to a collision. When the relative distance is zero, the relative speed is not zero. Generally, multiple spacecraft will intercept our spacecraft.

[0069] (2) There are 7 types of operational track intentions:

[0070] 1) Capture: The orbital movement distance is very short, usually intersecting, and the target is equipped with a robotic arm, grappling hook, etc.

[0071] 2) Taking photos: The relative distance is close enough, the lighting conditions are good, and the target has a CCD camera;

[0072] 3) Attack: causing physical damage to our spacecraft, such as impacting with a kinetic energy projectile, self-destructing impact of a low-value target, damaging our equipment with laser weapons, or bombarding our spacecraft with high-energy particles to cause malfunctions in our spacecraft's electronic equipment, etc.

[0073] 4) Electromagnetic interference: The target is relatively large and close to us, and is equipped with electromagnetic interference weapons, which affects our communications.

[0074] 5) Maintenance: Generally, the target is a cooperative objective, which includes maintenance tools such as robotic arms and equipment parts that need to be replaced or added. We support on-orbit maintenance.

[0075] 6) Refueling: This is generally for cooperative targets that are equipped with fuel packs, and we will support refueling.

[0076] 7) Electromagnetic attitude adjustment: When the relative distance is close, the target spacecraft carries a superconducting electromagnetic coil, which can change our attitude.

[0077] (3) There are 12 types of mission intents, among which some orbital or operational intents can be directly mission intents:

[0078] 1) Reconnaissance: The intention of the orbital movement is to maintain, escort, or fly around the target; the operational intention is to take pictures and obtain intelligence information from our side.

[0079] 2) Strikes: Partial disablement, with little impact on the on-orbit mission capability of our spacecraft; Destruction, with our spacecraft completely losing its on-orbit mission capability.

[0080] 3) Interference: The intended orbital maneuvers are escort / circumnavigation, interception, rendezvous, and expulsion; the operational intentions include electromagnetic interference, attack, and electromagnetic attitude adjustment. For a period of time, our side loses its on-orbit mission capabilities. Examples include electromagnetic interference, single-event inversion of electronic devices, and particle locking caused by high-energy particle bombardment.

[0081] 4) Position grabbing: The orbital motion intention is arbitrary, without any operational intent. The target anticipates the orbital intention of our spacecraft and maneuvers ahead of time to occupy the orbital position;

[0082] 5) Interception: The orbital maneuver is intended for interception, not for maneuvering. The target spacecraft forces our spacecraft to refrain from performing its planned orbital maneuvers.

[0083] 6) Protection: Cooperative targets protect our spacecraft and prevent non-cooperative targets from achieving their intentions to attack, seize position, intercept, or drive away our spacecraft;

[0084] 7) Monitoring: Generally for cooperative targets, the intentions of orbital motion and operation are similar to those of reconnaissance, which facilitates subsequent operations such as unwinding and maintenance;

[0085] 8) Departure: The orbital motion is intended to drive away the target spacecraft, not to manipulate it. The target spacecraft forces our spacecraft to leave its existing orbit.

[0086] 9) Negotiation: The intention of the orbital movement is to maintain and escort the aircraft around it, without any operational intention, and to conduct electromagnetic communication with our side;

[0087] 10) Standby: Similar to reconnaissance in terms of orbital movement intent, but without operational intent, waiting to receive instructions for operation;

[0088] 11) Maintenance: Generally a cooperative objective; the intention of the orbital movement is to rendezvous, and the intention of the operation is maintenance.

[0089] 12) Scientific experiments: Generally, these are cooperative objectives, carrying scientific experimental equipment. The intention of the orbital motion is to maintain, accompany / fly around, or rendezvous, and the operational intention is uncertain.

[0090] S2: As Figure 5 As shown, we analyze the information required to obtain the intent in typical scenarios and construct prompt elements for interaction with the large language model based on the required information.

[0091] S21: Obtain the possible information needed for the corresponding intent by judging typical scenarios;

[0092] S22: Based on the possible information required for the corresponding intention, summarize and classify the prompting elements, specifically five categories of prompting elements: situational situation, our information, target information, environmental conditions, and relative orbital motion characteristics;

[0093] S23: Expand the prompting elements of S22 with specific prompting words based on the required information in S21;

[0094] S24: The accuracy of the suggested words obtained from S23 is judged by experts;

[0095] S25: Organize the prompts obtained in S24 and generate, for example... Figure 6 The indicated keyword corpus.

[0096] By integrating and analyzing all possible information that could reveal the intentions of the target spacecraft in typical scenarios, five categories of clue elements are obtained: situational awareness, our own information, target information, environmental conditions, and relative orbital motion characteristics. Our own information specifically includes our purpose, our nature, our status, and our orbital characteristics. Sometimes, our purpose and orbital characteristics, as clue elements, can be redundant; that is, our purpose can be inferred from our orbital characteristics and other conditions, and vice versa. Therefore, only one of these clue elements needs to be provided. Target information includes target nature, target quantity, and target shape. Environmental conditions mainly include lighting conditions and electromagnetic conditions. The specific content includes the following 11 types:

[0097] (1) The situation refers to the state of affairs, which is divided into wartime and peacetime.

[0098] (2) Our purpose refers to the on-orbit missions of spacecraft, mainly including communication, data relay, navigation, meteorological observation, reconnaissance, resource exploration, and scientific exploration.

[0099] (3) Our side is divided into civilian and military. Non-cooperative targets have a stronger offensive intent towards our military satellites than civilian ones.

[0100] (4) Our status is divided into normal and faulty. The cooperation objective is more focused on repairing our faulty satellites.

[0101] (5) Our orbital characteristics can be divided into the following 8 types according to different orbital altitudes. The purpose of our satellite can be determined based on the orbital characteristics.

[0102] 1) Low Earth orbit (LEO): This refers to a near-circular orbit with an altitude below 2000km. Spacecraft are subject to weak atmospheric drag and need to maneuver to maintain their orbit. Typical spacecraft include manned spacecraft, space stations, Earth observation satellites, and some new types of communication satellites.

[0103] 2) Sun-synchronous orbit: The orbital altitude is approximately 600 to 800 km, the period is 96 to 100 minutes, the orbital inclination is approximately 98°, and the orbital plane precesses from west to east, with an average precession of 0.9856° per day. Typical spacecraft include meteorological satellites and remote sensing satellites.

[0104] 3) Medium Earth orbit (MEO): The orbital altitude is between 2,000 km and 35,786 km. Typical spacecraft are navigation satellites. The US GPS system, the Russian GLONASS system, the EU Galileo system, and the Chinese Beidou system all use a medium Earth orbit with an inclination of about 55° at an altitude of 20,000 to 24,000 km.

[0105] 4) Highly elliptical orbit (HEO): A highly elliptical orbit has a relatively low perigee and a very high apogee. Its apogee altitude is greater than 35,786 km, and its coverage time of the ground area below the apogee can exceed 12 hours. The most famous example is the Molniya orbit. Typical spacecraft using this orbit are communication satellites.

[0106] 5) Transfer orbit: This is a transitional orbit where the spacecraft stays for a very short time. Any attempt to influence the transfer orbit is aggressive and directly disrupts or delays the completion of our mission.

[0107] There are three types of transfer orbits: Medium Earth Transfer Orbit (MTO), with a perigee typically below 1000 km and an apogee between 2000 km and 35786 km. Generally, there are no specific restrictions on the perigee altitude of a MTO, but it usually does not exceed 400 km to reduce fuel requirements for orbital maneuvers. The resulting orbit is a Medium Earth orbit. Geostationary Transfer Orbit (GTO), with a perigee typically below 1000 km and an apogee of 35786 km, results in a Geostationary orbit. Super-Synchronous Transfer Orbit (SSTO), a special type of Geostationary Transfer Orbit with an apogee much greater than 35786 km, uses a double-elliptical pulse maneuver to reach Geostationary orbit; the first transfer orbit in this process is the Super-Synchronous Transfer Orbit.

[0108] 6) Geosynchronous orbit (GSO): The orbital period is equal to the Earth's rotation period, which is 1 sidereal day (23 hours, 56 minutes, and 4 seconds). The orbital altitude is 35,786 km, and the orbital inclination i < 90°. When the orbital inclination i = 0°, the relative position of the satellite and the ground remains unchanged, which is called a geostationary orbit.

[0109] 7) Inclined-Geosynchronous orbit: abbreviated as IGSO, is a geosynchronous orbit with an inclination of i≠0°. Its nadir trajectory is a closed curve in the shape of an "8". Three of the satellites of China's BeiDou Navigation Satellite System are located in an inclined geosynchronous orbit with an inclination of 55°.

[0110] 8) Geostationary orbit (GEO): This is a geosynchronous orbit with an inclination of i = 0°, where the satellite's relative position to the ground remains constant. It is primarily used for civilian and military satellites for communication, Earth observation, navigation, early warning, and meteorology.

[0111] (6) The nature of the goal is divided into cooperative goals and non-cooperative goals.

[0112] (7) The number of targets is divided into one and multiple. Multiple non-targets have a higher success rate in intercepting and driving away our side. Moreover, multiple small non-cooperative targets have a higher probability of attacking our side.

[0113] (8) Target shapes can be classified into small satellites and large satellites according to their size. Small satellites can be satellites equipped with cameras for reconnaissance missions or suicide satellites for attacking people. Large satellites, as long as they carry the corresponding equipment and instruments, can perform all of the above mission objectives. According to the specific details of their shape, they can be divided into operational shapes, such as those with robotic arms, grappling hooks, nets, electromagnetic coils, etc.; attack shapes, such as those with kinetic energy projectiles, laser weapons, high-energy particle emitters, etc.; communication shapes, such as those with antennas; and reconnaissance shapes, such as those with cameras.

[0114] (9) Lighting conditions are divided into facing the light and backlighting. Photo reconnaissance purposes require facing the light.

[0115] (10) Electromagnetic conditions are divided into good electromagnetic conditions and abnormal electromagnetic conditions. Good electromagnetic conditions can guarantee normal electromagnetic communication. Abnormal electromagnetic conditions, such as being located in the Van Allen radiation belts, cannot guarantee electromagnetic communication.

[0116] (11) Relative orbital characteristics are classified into the following four types based on relative distance and time:

[0117] 1) Flying past: The target spacecraft continuously approaches our side, reaches its closest distance, and then gradually moves away. This allows for observation of our spacecraft within a short period of time.

[0118] 2) Flying around or escorting: The target spacecraft is relatively close to our spacecraft and remains relatively constant, allowing for long-term observation of our spacecraft from different angles. Flying around at a sufficiently close distance can be achieved using grappling hooks, nets, electromagnetic coils, etc.

[0119] 3) Impact: The relative distance between the target spacecraft and our spacecraft eventually becomes 0, but the relative velocity is not 0, posing a threat to our spacecraft through a direct collision.

[0120] 4) Rendezvous: The relative distance between the target spacecraft and our spacecraft eventually becomes zero, enabling close-range, long-duration operations on our spacecraft. Suitable for missions with long operation times.

[0121] S3: Construct standard templates for prompt statements for large language model input;

[0122] Based on the five types of prompting elements and imperative prompting statements, a standard template for prompting statements in the input of a large language model is constructed, the specific content of which is as follows:

[0123] Standard phrase for describing the current situation: "Our spacecraft is currently in a state of ***."

[0124] Our standard statement: "Our spacecraft is a *** satellite, in *** orbit, and its status is ***."

[0125] Standard statement for target information: "The target spacecraft is *** target spacecraft, *** satellite, carrying ***."

[0126] Standard statement for environmental conditions: "Electromagnetic conditions ***, Illumination conditions ***".

[0127] The standard statement regarding relative orbital motion characteristics is: "The relative orbital motion characteristics of the target spacecraft and our spacecraft are ***."

[0128] The standard instruction statements may vary depending on the mission requirements. Here are two examples: "What is the intention of the target spacecraft? The answer should only provide the one most likely word of the intention, without any other unnecessary prompts" or "What is the intention of the target spacecraft? The answer should only provide the four most likely words of the intention, without any other unnecessary prompts."

[0129] S4: Construct sample data for large language model prompts based on prompt elements and standard templates for prompt statements;

[0130] The prompt words obtained in S2 are filled into the standard prompt statement template constructed in S3. After filtering and evaluating possible intentions, sample data is obtained. A specific example is as follows:

[0131] {"test":1,"prompt":["Our spacecraft is in wartime status","Our spacecraft is a military satellite in a sun-synchronous orbit, and its own status is normal.","The target spacecraft is a non-cooperative target spacecraft, a large satellite, equipped with a robotic arm.","Electromagnetic conditions are normal, illumination conditions are towards light.","The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous.","The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous."],"answer":["Capture","Strike","Interference"]}.

[0132] The answer is an adaptation to the probabilistic prediction characteristics of a large language model, and the number of answers is greater than 1.

[0133] The sample data obtained from S4 can be used directly for testing large language models without prior training.

[0134] The specific large language model chosen was glm-4, but its accuracy was too low in the test results to be directly applied. Therefore, the ChatGLM3 large language model was fine-tuned to implement the large language model for spacecraft intent recognition.

[0135] S5: Test the output of the large language model based on the sample data. The sample data obtained in S4 is used directly for testing the large language model without prior training. The specific large language model used is glm-4. Based on the test results, manual analysis shows that the large language model can be used in the field of spacecraft intent recognition. The criterion for judging the test results is that if the large language model outputs the expected intent, it is considered correct; otherwise, it is considered incorrect.

[0136] S6: After fine-tuning the training of the large language model, sample data testing is performed. The selected large language model is ChatGLM3. Several samples are randomly selected from the sample data obtained in S4 as training data, and the format of these training data is fine-tuned to meet the requirements of the command fine-tuning data test. Based on the large language model trained with fine-tuned commands, the intention of the target spacecraft is identified.

[0137] In summary, this invention provides a novel solution for spacecraft intent recognition, requiring only a small number of training samples compared to traditional learning methods. This invention addresses the current problem of incomplete spacecraft intent definitions and insufficient information utilization in this field.

[0138] The present invention will be further explained below with reference to specific embodiments.

[0139] Example

[0140] like Figure 2 As shown, the present invention includes the following steps:

[0141] S1: As Figure 3 As shown, a target spacecraft intent library is constructed for typical scenarios;

[0142] S11: Different intent attributes in typical scenarios are divided into three categories: track intent, operation intent, and task intent;

[0143] S12: Generate intent vocabulary under different intent attributes based on different intent attributes in typical scenarios;

[0144] S13: It is also possible to generate intent vocabulary under different intent attributes through a large language model;

[0145] S14: To ensure the accuracy of intent words, all intent words generated in S12 and S13 need to be analyzed and judged multiple times;

[0146] S15: Organize the intentional words correctly identified in S14, and obtain the following: Figure 4 The spacecraft shown is intended for use in a cargo warehouse.

[0147] S2: As Figure 5 As shown, we analyze the information required to obtain the intent in typical scenarios and construct prompt elements for interaction with the large language model based on the required information.

[0148] S21: Obtain the possible information needed for the corresponding intent by judging typical scenarios;

[0149] S22: Based on the possible information required for the corresponding intention, summarize and classify the prompting elements, specifically five categories of prompting elements: situational situation, our information, target information, environmental conditions, and relative orbital motion characteristics;

[0150] S23: Expand the prompting elements of S22 with specific prompting words based on the required information in S21;

[0151] S24: Determine the accuracy of the prompt words obtained from S23;

[0152] S25: Organize the prompts obtained in S24 and generate, for example... Figure 6 The indicated keyword corpus.

[0153] S3: Constructing a standard template for prompt statements for the input of a large language model; To ensure the accuracy and comprehensiveness of the semantics of the large language model input, and considering that the output format of the large language model should reflect spacecraft intent, a standard template for prompt statements for the input of a large language model is constructed based on five types of prompting elements and instructive prompt statements. The specific content is as follows:

[0154] Standard phrase for describing the current situation: "Our spacecraft is currently in a state of ***."

[0155] Our standard statement: "Our spacecraft is a *** satellite, in *** orbit, and its status is ***."

[0156] Standard statement for target information: "The target spacecraft is *** target spacecraft, *** satellite, carrying ***."

[0157] Standard statement for environmental conditions: "Electromagnetic conditions ***, Illumination conditions ***".

[0158] The standard statement regarding relative orbital motion characteristics is: "The relative orbital motion characteristics of the target spacecraft and our spacecraft are ***."

[0159] The standard instruction statements may vary depending on the mission requirements. Here are two examples: "What is the intention of the target spacecraft? The answer should only provide the one most likely word of the intention, without any other unnecessary prompts" or "What is the intention of the target spacecraft? The answer should only provide the four most likely words of the intention, without any other unnecessary prompts."

[0160] S4: Construct sample data for large language model prompts based on prompt elements and standard prompt statement templates; fill the prompt words obtained in S2 into the standard prompt statement templates constructed in S3 to obtain prompt statements; filter and evaluate possible intentions from S1 to obtain expected intentions; the prompt statements and expected intentions together constitute the sample data, and a total of 3048 test data entries are constructed. A specific example is as follows:

[0161] {"test":1,"prompt":["Our spacecraft is in wartime status","Our spacecraft is a military satellite in a sun-synchronous orbit, and its own status is normal.","The target spacecraft is a non-cooperative target spacecraft, a large satellite, equipped with a robotic arm.","Electromagnetic conditions are normal, illumination conditions are towards light.","The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous.","The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous."],"answer":["Capture","Strike","Interference"]}.

[0162] The answer is an adaptation to the probabilistic prediction characteristics of a large language model, and the number of answers is greater than 1.

[0163] S5: Test the output of the large language model based on sample data. The sample data obtained in S4 is used directly for testing the large language model without prior training. The specific large language model used is glm-4. Based on the test results, manual analysis shows that the large language model can be used in the field of spacecraft intent recognition. The criterion for judging the test results is that the large language model outputs the expected intent and is considered correct; otherwise, it is considered incorrect. Statistically, the accuracy rate of the S3 standard instruction prompt statement "What is the intent of the target spacecraft? The answer should only give the one most likely word of the intent, without any other unnecessary prompts" is 4.46%, and the accuracy rate of the standard instruction prompt statement "What is the intent of the target spacecraft? The answer should only give the four most likely words of the intent, without any other unnecessary prompts" is 41.08%.

[0164] S6: After fine-tuning the training of the large language model's instructions, test it with sample data. The large language model used is ChatGLM3. Eighty samples were randomly selected from the sample data obtained in S4 as training data. The format of these 80 training data samples was fine-tuned to meet the requirements of the instruction fine-tuning data test. The specific training data format is as follows:

[0165] {

[0166] "instruction": "Based on the following prompts, select the possible intentions of the target spacecraft from 'capture, photograph, refuel, electromagnetic attitude adjustment, electromagnetic interference, maintenance, reconnaissance, monitoring, strike, expel, interfere, negotiate, standby, seize position, maintenance, intercept, escort, scientific experiment'. The answer should only provide the possible intentions and should not contain any other words."

[0167] "input": "Our spacecraft is currently in a wartime state. Our spacecraft is a military satellite in a sun-synchronous orbit, and its own condition is normal. The target spacecraft is a non-cooperative target spacecraft, a large satellite equipped with a robotic arm. Electromagnetic conditions are normal, and illumination conditions are in the direction of light. The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous."

[0168] "output": "Capture, strike, interfere"

[0169] }

[0170] We then trained ChatGLM3 with instruction fine-tuning. This resulted in the ChatGLM3 instruction fine-tuning model. The basic ChatGLM3 model and the instruction fine-tuning model were merged, and the same evaluation criteria as S5 were used to test the data. The test data was then changed to the following format:

[0171] {

[0172] "test":1,

[0173] "prompt":["Instruction: Based on the following prompts, select the possible intentions of the target spacecraft from "capture, photograph, refuel, electromagnetic attitude adjustment, electromagnetic interference, maintenance, reconnaissance, monitoring, strike, expulsion, interference, negotiation, standby, positioning, maintenance, interception, escort, scientific experiment". Only provide the possible intentions; no other unnecessary words are needed.\nInput: Our spacecraft is in a wartime state. Our spacecraft is a military satellite in a sun-synchronous orbit, and its own condition is normal. The target spacecraft is a non-cooperative target spacecraft, a large satellite, equipped with a robotic arm. Electromagnetic conditions are normal, and illumination conditions are towards the sun. The relative orbital motion characteristics of the target spacecraft and our spacecraft are rendezvous."]

[0174] "answer":["arrest","strike","interference"]

[0175] }

[0176] Statistical analysis shows that the accuracy of the fine-tuned model is 99.51%, making it suitable for spacecraft intent recognition tasks.

[0177] The second objective of this invention is to propose a target spacecraft intent recognition system based on a large language model, comprising:

[0178] The intent library building module is used to build the intent library for the target spacecraft.

[0179] Corpus building module: used to build prompt elements that interact with the large language model based on the target spacecraft intent library, and generate a prompt word corpus;

[0180] Training Model Module: Used to train the large language model based on the prompt word corpus to obtain the trained large language model;

[0181] Intent recognition module: Used to identify the intent of the target spacecraft using a trained large language model.

[0182] A third objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the target spacecraft intent recognition method based on a large language model.

[0183] The target spacecraft intent recognition method based on a large language model includes the following steps:

[0184] Build a target spacecraft intent database;

[0185] Based on the target spacecraft intent database, construct prompting elements that interact with a large language model and generate a prompting word corpus;

[0186] Based on the prompt word corpus, the large language model is trained to obtain the trained large language model;

[0187] The trained large language model is used to identify the target spacecraft's intent.

[0188] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the target spacecraft intent recognition method based on a large language model.

[0189] The target spacecraft intent recognition method based on a large language model includes the following steps:

[0190] Build a target spacecraft intent database;

[0191] Based on the target spacecraft intent database, construct prompting elements that interact with a large language model and generate a prompting word corpus;

[0192] Based on the prompt word corpus, the large language model is trained to obtain the trained large language model;

[0193] The trained large language model is used to identify the target spacecraft's intent.

[0194] A fifth objective of this invention is to provide a computer program product comprising computer instructions that, when executed by a processor, implement the above-described... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0195] Many embodiments and applications beyond the examples provided will be apparent to those skilled in the art upon reading the foregoing description. Therefore, the scope of this teaching should not be determined by reference to the foregoing description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed inventive subject matter.

[0196] The above content provides a further detailed description of the present invention. It should not be construed that the specific embodiments of the present invention are limited to this. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention as defined by the submitted claims.

Claims

1. A method for recognizing the intent of a target spacecraft based on a large language model, characterized in that, include: Build a target spacecraft intent database; Based on the target spacecraft intent database, construct prompting elements that interact with a large language model and generate a prompting word corpus; Based on the prompt word corpus, the large language model is trained to obtain the trained large language model; The trained large language model is used to identify the target spacecraft's intent; The spacecraft intent library includes orbital intents, operational intents, and mission intents; The process involves constructing prompting elements based on the target spacecraft's intent database and interacting with a large language model to generate a prompting word corpus, including: Based on the target spacecraft intent database, determine the information needed to obtain the corresponding intent in typical scenarios; Based on the information needed to obtain the corresponding intent, construct prompt elements for interaction with the large language model; The prompting elements for interaction with the large language model are expanded to obtain prompting vocabulary; Organize the obtained prompt vocabulary and generate a prompt corpus; The prompting elements for interaction with the large language model include the current situation, our information, target information, environmental conditions, and relative orbital motion characteristics.

2. The target spacecraft intent recognition method based on a large language model according to claim 1, characterized in that, The construction of the target spacecraft intent library involves generating intent words under different intent attributes using a large language model, then analyzing and judging all the generated intent words, and organizing the correctly judged intent words to obtain the target spacecraft intent library.

3. The target spacecraft intent recognition method based on a large language model according to claim 1, characterized in that, The step of training a large language model based on a prompt word corpus to obtain a trained large language model includes: Construct standard templates for prompt statements for input to a large language model based on a corpus of prompt words; Sample data for input prompts are constructed based on prompt elements and standard templates for prompt statements in a large language model. The large language model is trained by fine-tuning the input sample data prompted by the large language model, and the trained large language model is obtained.

4. The target spacecraft intent recognition method based on a large language model according to claim 3, characterized in that, The process of constructing sample data for prompt input based on prompt elements and prompt statement standard templates involves filling prompt words into prompt statement standard templates, filtering and evaluating possible intentions to obtain sample data for prompt input from the large language model.

5. A target spacecraft intent recognition system based on a large language model, comprising the target spacecraft intent recognition method based on a large language model as described in any one of claims 1 to 4, characterized in that, include: The intent library building module is used to build the intent library for the target spacecraft. Corpus building module: used to build prompt elements that interact with the large language model based on the target spacecraft intent library, and generate a prompt word corpus; Training Model Module: Used to train the large language model based on the prompt word corpus to obtain the trained large language model; Intent recognition module: Used to identify the intent of the target spacecraft using a trained large language model.

6. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the target spacecraft intent recognition method based on any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the target spacecraft intent recognition method based on a large language model as described in any one of claims 1-4.

Images