Target satellite group intention recognition method based on unsupervised learning

Abstract

The invention discloses a target satellite group intention recognition method based on unsupervised learning, and belongs to the technical field of space safety and maintenance, and the method comprises the following steps: constructing a target satellite group task intention set; building a convolutional neural network based on a TS2Vec self-supervised learning framework, and training the neural network to learn three-dimensional lattice sequence features of target star group configuration; based on a DBSCAN density clustering algorithm, grouping target star groups to configure a three-dimensional lattice sequence; and applying the trained algorithm to target satellite group intention recognition. According to the method, the intention recognition problems that an existing algorithm is difficult to process, label sample data is scarce, data noise is strong, the nature of the configuration is difficult to analyze, and unknown intentions emerge continuously and are difficult to extend are solved.

Description

Technical Field

[0001] This invention belongs to the field of space security and maintenance technology, specifically relating to a target constellation intent recognition method based on unsupervised learning. Background Technology

[0002] With the rapid development of aerospace technology and the widespread adoption of low-cost satellite launch technology, the number of spacecraft in low Earth orbit and deep space exploration missions has increased dramatically. The scale and complexity of space targets, especially constellations and clusters of multiple satellites, are also increasing. Against this backdrop, timely and accurate identification of the behavior patterns and mission intentions of target constellations has become a core requirement in fields such as space situational awareness, space security early warning, and on-orbit mission planning.

[0003] Traditional methods for identifying the intent of space targets primarily rely on prior knowledge bases and rule-based reasoning. Specifically, they typically establish a behavioral rule base based on known target orbital parameters, platform characteristics, and mission history data, combined with expert experience. The target's intent is then inferred by matching observational data with pre-defined rules. These methods are effective when target behavior rules are clear and prior information is complete. However, facing the increasing number of novel, intelligent, or adaptive constellation systems, whose behavior patterns may dynamically change and whose collaborative strategies are complex and obscure, traditional methods have significant limitations: First, they heavily rely on manually defined rules, making it difficult to cover unknown or sudden behavioral patterns; second, for constellations with autonomous collaborative capabilities, their collective intent often exhibits distributed and emergent characteristics, making it difficult to summarize using simple rules; third, space observation data often suffers from high noise, incompleteness, and discontinuity, resulting in insufficient robustness of systems based on deterministic rules. In recent years, data-driven constellation behavior analysis has largely focused on single-target trajectory prediction or simple cluster analysis, failing to delve into the complex spatiotemporal relationships among multiple targets within a constellation, their collaborative motion patterns, and their intrinsic connection to higher-level strategic intent. Unsupervised learning techniques offer a new approach to solving the aforementioned problems due to their ability to discover the inherent structure and distribution of data without requiring labeled data.

[0004] Therefore, in view of the problems of existing technologies such as strong dependence on prior knowledge and labeled data, weak ability to recognize unknown behavior patterns, and insufficient mining of star cluster collaborative intentions, there is an urgent need for a target star cluster intention recognition method based on unsupervised learning. Summary of the Invention

[0005] This invention provides a target constellation intent recognition method based on unsupervised learning, which solves the intent recognition problems that existing algorithms struggle with, such as scarce training labeled sample data, strong data noise making it difficult to capture the essence of the configuration, and the continuous emergence of unknown intents that are difficult to extrapolate.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is: a target star cluster intent recognition method based on unsupervised learning, comprising the following steps: S1. Construct the target constellation mission intent set; S2. Based on the TS2Vec self-supervised learning framework, a convolutional neural network is built and trained to learn the three-dimensional lattice sequence features of the target star cluster configuration. S3. Based on the DBSCAN density clustering algorithm, group the three-dimensional lattice sequence of target star cluster configuration; S4. Apply the trained algorithm to target constellation intent recognition.

[0007] Furthermore, the specific steps of step S1 are as follows: S11. Based on the payload carried by the satellite, simplify the satellite types in the target constellation to main satellite, communication satellite, observation satellite, and functional satellite; S12. Based on the functional division and motion characteristics of satellites, the complex satellite constellation mission intent is abstracted into ten mission intents: encirclement, reconnaissance, tracking, interference, assembly, dispersion, evasion, escape, protection, and standby. S13. Based on the definition of the target constellation mission intent, construct a three-dimensional lattice sequence example of the target constellation mission intent configuration.

[0008] Furthermore, the satellite types in the target constellation in step S11 are defined as follows: The primary satellite holds a central leadership position in the target constellation, responsible for collecting and analyzing environmental awareness information from all other satellites in the constellation, making decisions and allocating tasks for space missions, and directing the maneuvers, grouping, and space operations of communication satellites, observation satellites, and functional satellites. Communication satellites are responsible for enabling information exchange between the main satellite and communication satellites, between the main satellite and functional satellites, and between observation satellites and functional satellites. When the distance between the main satellite and the observation satellite or functional satellite exceeds the maximum communication distance, or when the number of communications reaches the maximum limit, communication satellites are needed to enable information exchange among satellites in the constellation. The observation satellite is responsible for acquiring the status information of space targets and other environmental information, and uploading the acquired observation information to the main satellite for decision-making. Functional satellites are satellites that actually perform space missions. They carry various payloads required for the mission, such as docking devices, capture devices, reconnaissance devices, and electromagnetic interference devices. They are commanded by the host satellite and perform maneuvers, formations, and space operations. The target constellation mission intent in step S12 is defined as follows: Encirclement refers to a state in which a certain number of target constellation satellites approach our spacecraft from multiple directions and multiple orbital planes, forming a geometric distribution and motion pattern in the space region around our spacecraft. Reconnaissance refers to the state in which a certain number of target satellite constellations pass through the space region surrounding our spacecraft in a relatively flying orbit. Tracking refers to the state in which a certain number of target satellite constellations maintain a certain distance from our spacecraft by taking relative motion orbits, either around or accompanying it. Interference refers to the state in which a certain number of target satellite constellations repeatedly pass through the space region near the transfer orbit of our spacecraft in a short period of time, adopting a relatively fast-moving, skimming trajectory. Gathering refers to the state in which a group of satellites approach a certain location in space by taking a relative motion orbit, either around or accompanying it, and the relative distance between them is no greater than a certain value within a certain period of time. Dispersion refers to a situation where the relative distances between satellites in a target constellation change from a state where the distances remain constant for a long period of time to a state where they generally increase. Evasion refers to the state of actively avoiding a certain space region during the transfer of a target star cluster to a certain location; Escape refers to the state in which the distance between the target constellation satellites and our spacecraft gradually increases; "Escort" refers to the state in which most of the satellites in a target constellation move in relative orbits or fly alongside one or more of the satellites in the target constellation. Standby refers to the state in which the satellites of a target constellation maintain their current relative orbit for an extended period of time; The target constellation configuration three-dimensional lattice sequence in step S13 refers to mapping the position of each satellite in the target constellation in the LVLH coordinate system onto the target constellation configuration three-dimensional lattice map, and then taking three-dimensional lattice maps at equal time intervals to form the target constellation configuration three-dimensional lattice sequence, the data shape of which is as follows: N × x × y × z × s The specific meaning of each dimension of data is as follows: N This refers to the first [number] point in the three-dimensional lattice sequence of the target star cluster configuration. N The frame contains a 3D dot matrix of the target star cluster configuration, with values ​​ranging from 0 to 30. The time interval between adjacent 3D dot matrix images of the target star cluster configuration is 1 minute. x This refers to the satellite's position in the LVLH coordinate system. x The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. x The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. x =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. y This refers to the satellite's position in the LVLH coordinate system. yThe axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. y The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. y =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. z This refers to the satellite's position in the LVLH coordinate system. z The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. z The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. z =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. s This refers to the type of satellite located at this coordinate on the three-dimensional dot matrix diagram of the target constellation configuration. Different values ​​represent: 0 indicates no satellite at this coordinate, 1 indicates the target constellation's main satellite at this coordinate, 2 indicates the target constellation's communication satellite at this coordinate, 3 indicates the target constellation's observation satellite at this coordinate, 4 indicates the target constellation's functional satellite at this coordinate, and 5 indicates our spacecraft at this coordinate.

[0009] Furthermore, the specific steps of step S2 are as follows: S21. Augment the current batch of target star cluster configuration 3D lattice sequence by rotation, translation and cropping to generate positive and negative sample pairs; S22. Construct a convolutional neural network feature extractor. Input the augmented 3D point matrix sequence into the convolutional neural network feature extractor to extract the feature vector of the 3D point matrix sequence. S23. Construct a multilayer perceptron neural network projector, input the feature vector into the multilayer perceptron neural network projector, and output the projection vector. S24. Design a contrastive loss function and train a convolutional neural network based on the projection vector to learn the general features of the three-dimensional lattice sequence of the target star cluster configuration.

[0010] Furthermore, in step S21, data augmentation and positive / negative sample pair generation refer to applying a set of random but physically meaningful transformations to generate a "view" of a batch of input 3D point matrix sequence samples, forming positive sample pairs. These transformations aim to simulate the uncertainty in observation while maintaining the intended semantics. The specific method of data augmentation is as follows: Spatial rotation refers to a small, random rotation around a certain axis in the LVLH coordinate system to simulate a minute change in the observation perspective; Random translation refers to a small-scale random translation within the entire lattice space; Sequence pruning refers to the random selection of continuous segments in the time dimension; The convolutional neural network feature extractor in step S22 is based on a convolutional neural network design, with the input layer being... N × x × y × z × s The target star cluster configuration three-dimensional point array sequence data is processed through five convolutional layers for adaptive pooling, and then through three fully connected layers to finally output the feature vector of the three-dimensional point array sequence. The multilayer perceptron neural network projector in step S23 is composed of two layers of multilayer perceptron neural network. It maps the feature vectors output by the feature extractor to the contrast learning space, so that the projection vectors of similar images are closer together on the unit hypersphere, providing an optimization target for loss calculation. The expression for the contrast loss function in step S24 is as follows: in, z i and z j These represent two feature vectors of the same positive sample pair. z k The feature vector representing the negative sample. This indicates that the cosine similarity is taken. Indicates temperature hyperparameter, 2 N This indicates the number of data items in a batch.

[0011] Furthermore, the specific steps of step S3 are as follows: S31. Based on experience and the elbow rule, set the initial parameters of the DBSCAN density clustering algorithm; S32. Input the feature vectors extracted from step S2 into the DBSCAN density clustering algorithm. Based on the local density of the input dataset, discover intention clusters of arbitrary shape and number, and adaptively adjust the algorithm parameters. S33. The clustering results aggregate the features of the three-dimensional lattice sequence of the target star cluster configuration into multiple intent clusters, and assign task intent semantic labels to the intent clusters. S34. Construct and maintain the target constellation mission intent feature set formed by clustering results.

[0012] Furthermore, in step S31, the initial parameters of the DBSCAN density clustering algorithm are the neighborhood radius. e The minimum number of neighborhood samples, MinPts, for the core points; The specific steps of step S32 are as follows: S321. Based on experience, set a smaller [value]. einitial value e 0 and a fixed MinPts value; S322, will e Inputting 0 as a candidate value into DBSCAN performs clustering to obtain the number of clusters. C 0; S323. Increase the candidate value by a fixed step size. s The number of clusters is obtained by clustering again. C 1; S324, Repeat step S323; S325. Monitor changes in the number of clusters; if a preset number is detected... n Continuous candidate values e k , e k+1 , ..., e n Number of clusters generated C k , C k+1 ,,…, C n If all candidate values ​​are identical, the clustering result is considered to have entered a relatively stable plateau region, and the largest value among these candidate values ​​is selected as the final result. e parameter; In step S33, assigning task intent semantic labels to intent clusters means that by associating them with the target constellation task intents defined in step S12, each stable cluster with a large amount of data can be assigned a corresponding semantic label. For feature vectors that cannot be classified into any existing cluster or a new, small and dense cluster, the system can identify them as potential unknown intents, trigger an alarm, and hand them over to experts for analysis. The construction and maintenance of the target constellation mission intent feature set formed by the clustering results in step S34 refers to the construction of the target constellation mission intent feature library formed by clustering training data, which includes the target constellation mission intents defined in step S12. As new data is accumulated and confirmed, potential mission intents can be added to the intent library to realize the incremental extensional learning of the model and expand the target constellation mission intent feature set.

[0013] Furthermore, the specific steps of step S4 are as follows: S41. Load the pre-trained convolutional neural network of the feature extractor and the target constellation mission intent feature library into the onboard computer of our spacecraft. S42. For the acquired target star cluster observation data, convert it into a three-dimensional lattice sequence of the target star cluster configuration according to the method described in step S13. S43. Input the three-dimensional lattice sequence of the target star cluster configuration into the feature extractor to extract the target star cluster configuration features; S44. Compare the feature vector with the constructed target constellation mission intent feature library, calculate its distance from the center of each intent cluster, and output the target constellation mission intent label and confidence level.

[0014] The beneficial effects of the present invention are: (1) A target star cluster intention recognition method is designed based on the unsupervised learning algorithm framework, which enables it to recognize the other party's intention according to the configuration changes of the target star cluster. The neural network is trained on the ground and then the trained neural network is loaded on the satellite, which reduces the computing power burden of the onboard computer and speeds up the intention recognition speed. (2) Based on the TS2Vec self-supervised learning framework, this invention builds a convolutional neural network to learn the three-dimensional lattice sequence features of the target star cluster configuration. It overcomes the problems of strong dependence on prior knowledge and labeled data in existing technologies and the fact that space observation data often has large noise, incompleteness and discontinuity. It gets rid of dependence on manually labeled data, suppresses intra-class noise, strengthens inter-class differences, and autonomously learns the general features of the target star cluster configuration. (3) This invention develops a DBSCAN density clustering algorithm and incremental mechanism to address the unknown intentions emerging from the target star cluster. It overcomes the difficulty of traditional methods in dealing with the distributed and emergent characteristics of group intentions, adaptively adjusts the algorithm parameters, autonomously mines the potential behavioral structure of the star cluster, and classifies and generates new intention clusters, thereby realizing the intention recognition of the target star cluster. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a target constellation intent recognition method based on unsupervised learning according to the present invention.

[0016] Figure 2 This is a flowchart illustrating the specific steps of step S1 in the target constellation intent recognition method based on unsupervised learning of the present invention.

[0017] Figure 3 This is a flowchart illustrating the specific steps of step S2 in the target constellation intent recognition method based on unsupervised learning of the present invention.

[0018] Figure 4 This is a flowchart illustrating the specific steps of step S3 in the target constellation intent recognition method based on unsupervised learning of the present invention.

[0019] Figure 5 This is a flowchart illustrating the specific steps of step S4 in the target constellation intent recognition method based on unsupervised learning of the present invention.

[0020] Figure 6 This is a schematic diagram illustrating the "encirclement" intent of a target star cluster in a target star cluster intent recognition method based on unsupervised learning according to the present invention.

[0021] Figure 7This is a schematic diagram illustrating the "reconnaissance" intent of a target star cluster in a target star cluster intent recognition method based on unsupervised learning according to the present invention.

[0022] Figure 8 This is a schematic diagram illustrating the "tracking" intent of a target constellation in a target constellation intent recognition method based on unsupervised learning according to the present invention.

[0023] Figure 9 This is a schematic diagram of the "interference" intent of a target star cluster in the target star cluster intent recognition method based on unsupervised learning of the present invention.

[0024] Figure 10 This is a schematic diagram of the "aggregation" intent of a target star cluster in the target star cluster intent recognition method based on unsupervised learning of the present invention.

[0025] Figure 11 This is a schematic diagram illustrating the "dispersed" intent of a target star cluster in a target star cluster intent recognition method based on unsupervised learning according to the present invention.

[0026] Figure 12 This is a schematic diagram illustrating the "avoidance" intent of a target star cluster in a target star cluster intent recognition method based on unsupervised learning according to the present invention.

[0027] Figure 13 This is a schematic diagram illustrating the "escape" intent of a target constellation in a target constellation intent recognition method based on unsupervised learning according to the present invention.

[0028] Figure 14 This is a schematic diagram of the "guardian" intent of a target star cluster in the target star cluster intent recognition method based on unsupervised learning of the present invention.

[0029] Figure 15 This is a schematic diagram of the "waiting" intent of a target star cluster in a target star cluster intent recognition method based on unsupervised learning according to the present invention. Detailed Implementation Plan Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

[0030] Example like Figure 1 As shown, this invention provides a target constellation intent recognition method based on unsupervised learning, comprising the following steps: S1. Construct the target constellation mission intent set; S2. Based on the TS2Vec self-supervised learning framework, a convolutional neural network is built and trained to learn the three-dimensional lattice sequence features of the target star cluster configuration. S3. Based on the DBSCAN density clustering algorithm, group the three-dimensional lattice sequence of target star cluster configuration; S4. Apply the trained algorithm to target constellation intent recognition.

[0031] The specific steps of step S1 are as follows: S11. Based on the payload carried by the satellite, simplify the satellite types in the target constellation to main satellite, communication satellite, observation satellite, and functional satellite; S12. Based on the functional division and motion characteristics of satellites, the complex satellite constellation mission intent is abstracted into ten mission intents: encirclement, reconnaissance, tracking, interference, assembly, dispersion, evasion, escape, protection, and standby. S13. Based on the definition of the target constellation mission intent, construct a three-dimensional lattice sequence example of the target constellation mission intent configuration.

[0032] The satellite type in the target constellation in step S11 is defined as follows: The primary satellite holds a central leadership position in the target constellation, responsible for collecting and analyzing environmental awareness information from all other satellites in the constellation, making decisions and allocating tasks for space missions, and directing the maneuvers, grouping, and space operations of communication satellites, observation satellites, and functional satellites. Communication satellites are responsible for enabling information exchange between the main satellite and communication satellites, between the main satellite and functional satellites, and between observation satellites and functional satellites. When the distance between the main satellite and the observation satellite or functional satellite exceeds the maximum communication distance, or when the number of communications reaches the maximum limit, communication satellites are needed to enable information exchange among satellites in the constellation. The observation satellite is responsible for acquiring the status information of space targets and other environmental information, and uploading the acquired observation information to the main satellite for decision-making. Functional satellites are satellites that actually perform space missions. They carry various payloads required for the mission, such as docking devices, capture devices, reconnaissance devices, and electromagnetic interference devices. They are commanded by the host satellite and perform maneuvers, formations, and space operations. The target constellation mission intent in step S12 is defined as follows: Encirclement refers to a state in which a certain number of target constellation satellites approach our spacecraft from multiple directions and multiple orbital planes, forming a geometric distribution and motion pattern in the space region around our spacecraft. Reconnaissance refers to the state in which a certain number of target satellite constellations pass through the space region surrounding our spacecraft in a relatively flying orbit. Tracking refers to the state in which a certain number of target satellite constellations maintain a certain distance from our spacecraft by taking relative motion orbits, either around or accompanying it. Interference refers to the state in which a certain number of target satellite constellations repeatedly pass through the space region near the transfer orbit of our spacecraft in a short period of time, adopting a relatively fast-moving, skimming trajectory. Gathering refers to the state in which a group of satellites approach a certain location in space by taking a relative motion orbit, either around or accompanying it, and the relative distance between them is no greater than a certain value within a certain period of time. Dispersion refers to a situation where the relative distances between satellites in a target constellation change from a state where the distances remain constant for a long period of time to a state where they generally increase. Evasion refers to the state of actively avoiding a certain space region during the transfer of a target star cluster to a certain location; Escape refers to the state in which the distance between the target constellation satellites and our spacecraft gradually increases; "Escort" refers to the state in which most of the satellites in a target constellation move in relative orbits or fly alongside one or more of the satellites in the target constellation. Standby refers to the state in which the satellites of a target constellation maintain their current relative orbit for an extended period of time; The target constellation configuration three-dimensional lattice sequence in step S13 refers to mapping the position of each satellite in the target constellation in the LVLH coordinate system onto the target constellation configuration three-dimensional lattice map, and then taking three-dimensional lattice maps at equal time intervals to form the target constellation configuration three-dimensional lattice sequence, the data shape of which is as follows: N × x × y × z × s The specific meaning of each dimension of data is as follows: N This refers to the first [number] point in the three-dimensional lattice sequence of the target star cluster configuration. N The frame contains a 3D dot matrix of the target star cluster configuration, with values ​​ranging from 0 to 30. The time interval between adjacent 3D dot matrix images of the target star cluster configuration is 1 minute. x This refers to the satellite's position in the LVLH coordinate system. x The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. x The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. x =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. y This refers to the satellite's position in the LVLH coordinate system. y The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. y The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. y=0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. z This refers to the satellite's position in the LVLH coordinate system. z The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. z The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. z =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. s This refers to the type of satellite located at this coordinate on the three-dimensional dot matrix diagram of the target constellation configuration. Different values ​​represent: 0 indicates no satellite at this coordinate, 1 indicates the target constellation's main satellite at this coordinate, 2 indicates the target constellation's communication satellite at this coordinate, 3 indicates the target constellation's observation satellite at this coordinate, 4 indicates the target constellation's functional satellite at this coordinate, and 5 indicates our spacecraft at this coordinate.

[0033] In this embodiment, the LVLH coordinate system is defined as follows: the origin of the coordinate system is the center of mass of the reference spacecraft. x The direction is from the origin to the center of mass. z The direction is perpendicular to the orbital plane and parallel to the direction of the angular momentum of the reference spacecraft orbiting the Earth. y The direction is the direction that makes the coordinate system form a right-handed rectangular coordinate system.

[0034] The reference spacecraft orbit is set near the target constellation orbit, centered on our spacecraft or a specific point of interest, so that the target constellation is located in a cubic space with the origin as the center, a side length of 12,800 meters and parallel to the coordinate axes in the LVLH coordinate system defined by the reference spacecraft.

[0035] Establish an LVLH coordinate system centered on our spacecraft or a specific point of interest, and define a three-dimensional spatial grid with its origin coinciding with the origin of the LVLH coordinate system. x , y , z The range of each of the three axes is from -128 to 127, totaling 256 graduations. The interval between adjacent graduations in real space is 50 meters. A cube space with a side length of 12.8 kilometers is discretized into a 256×256×256 lattice.

[0036] For a continuous observation period, sampling is performed at fixed intervals of 1 minute. At each sampling moment, the LVLH coordinates of each satellite in the constellation are mapped onto the aforementioned three-dimensional grid, and the nearest grid point is found. The value at that grid point is... sSatellites are labeled according to their roles: 1 (primary satellite), 2 (communication satellite), 3 (observation satellite), 4 (functional satellite), and 5 (our own spacecraft). If there is no satellite, it is labeled 0. This is based on 31 consecutive time points, spanning 30 minutes. N A raster graph with values ​​from 0 to 30, stacked in chronological order, forms a five-dimensional tensor with the following shape: N × x × y × z × s That is, 31×256×256×256×1.

[0037] This tensor fully encodes the spatiotemporal evolution information of the star cluster over half an hour, including the distribution of roles and relative motion trajectories, and serves as a direct input for subsequent feature learning.

[0038] The specific steps of step S2 are as follows: S21. Augment the current batch of target star cluster configuration 3D lattice sequence by rotation, translation and cropping to generate positive and negative sample pairs; S22. Construct a convolutional neural network feature extractor. Input the augmented 3D point matrix sequence into the convolutional neural network feature extractor to extract the feature vector of the 3D point matrix sequence. S23. Construct a multilayer perceptron neural network projector, input the feature vector into the multilayer perceptron neural network projector, and output the projection vector. S24. Design a contrastive loss function and train a convolutional neural network based on the projection vector to learn the general features of the three-dimensional lattice sequence of the target star cluster configuration.

[0039] The data augmentation and positive / negative sample pair generation in step S21 refers to applying a set of random but physically meaningful transformations to generate a "view" of a batch of 3D point matrix sequence samples to form positive sample pairs. These transformations aim to simulate the uncertainty in observation while maintaining the intended semantics. The specific methods of data augmentation are as follows: Spatial rotation refers to a small, random rotation around a certain axis in the LVLH coordinate system to simulate a minute change in the observation perspective; Random translation refers to a small-scale random translation within the entire lattice space; Sequence pruning refers to the random selection of continuous segments in the time dimension; In this embodiment, through the above transformation, two different views generated from the same original sequence are considered semantically consistent positive sample pairs. All other transformed sequence samples in the same batch are considered negative samples of that original sample.

[0040] In step S22, the convolutional neural network feature extractor is designed based on a convolutional neural network, with the input layer being...N × x × y × z × s The target star cluster configuration three-dimensional point array sequence data is processed through five convolutional layers for adaptive pooling, and then through three fully connected layers to finally output the feature vector of the three-dimensional point array sequence. In this embodiment, the input to the feature extractor is a five-dimensional tensor with a shape of 31×256×256×256×1. The main body of the network consists of five three-dimensional convolutional layers, each followed by a normalization and ReLU activation function, and an adaptive three-dimensional pooling layer is used to gradually compress the spatial dimension.

[0041] Convolutional layers are used to capture patterns of star cluster configurations in local spatial regions. The feature maps are flattened and passed through three fully connected layers, outputting a 1024-dimensional feature vector. h This vector is a high-level, dense representation of the input sequence.

[0042] In step S23, the multilayer perceptron neural network projector is composed of two layers of multilayer perceptron neural network. It maps the feature vectors output by the feature extractor to the contrast learning space, so that the projection vectors of similar images are closer together on the unit hypersphere, providing an optimization target for loss calculation. In this embodiment, the projection head consists of a two-layer multilayer perceptron neural network, and the input is a feature vector. h The output is a projection vector that is more suitable for contrastive learning. z This step maps the features onto a unit hypersphere, such that, after optimization with contrastive loss, the cosine distance between the projection vectors of positive sample pairs on this hypersphere is as small as possible (high similarity), while the cosine distance with negative sample pairs is as large as possible (low similarity).

[0043] The expression for the contrast loss function in step S24 is as follows: in, z i and z j This refers to two samples that represent the same pair of positive samples. z k The feature vector representing the negative sample. This indicates that the cosine similarity is taken. Indicates temperature hyperparameter, 2 N This indicates the number of data items in a batch.

[0044] In this embodiment, the feature extractor is optimized through training on large-scale unlabeled data to extract feature representations that remain unchanged when augmented with the above data and can distinguish different star cluster dynamic modes.

[0045] The specific steps of step S3 are as follows: S31. Based on experience and the elbow rule, set the initial parameters of the DBSCAN density clustering algorithm; S32. Input the feature vectors extracted from step S2 into the DBSCAN density clustering algorithm. Based on the local density of the input dataset, discover intention clusters of arbitrary shape and number, and adaptively adjust the algorithm parameters. S33. The clustering results aggregate the features of the three-dimensional lattice sequence of the target star cluster configuration into multiple intent clusters, and assign task intent semantic labels to the intent clusters. S34. Construct and maintain the target constellation mission intent feature set formed by clustering results.

[0046] In step S31, the initial parameters of the DBSCAN density clustering algorithm are the neighborhood radius. e The minimum number of neighborhood samples, MinPts, for the core points; The specific steps of step S32 are as follows: S321. Based on experience, set a smaller [value]. e initial value e 0 and a fixed MinPts value; S322, will e Inputting 0 as a candidate value into DBSCAN performs clustering to obtain the number of clusters. C 0; S323. Increase the candidate value by a fixed step size. s The number of clusters is obtained by clustering again. C 1; S324, Repeat step S323; S325. Monitor changes in the number of clusters; if a preset number is detected... n Continuous candidate values e k , e k+1 , ..., e n Number of clusters generated C k , C k+1 ,,…, C n If all candidate values ​​are identical, the clustering result is considered to have entered a relatively stable plateau region, and the largest value among these candidate values ​​is selected as the final result. e parameter; In step S33, assigning task intent semantic labels to intent clusters means that by associating them with the target constellation task intents defined in step S12, each stable cluster with a large amount of data can be assigned a corresponding semantic label. For feature vectors that cannot be classified into any existing cluster or a new, small and dense cluster, the system can identify them as potential unknown intents, trigger an alarm, and hand them over to experts for analysis. The construction and maintenance of the target constellation mission intent feature set formed by the clustering results in step S34 refers to the construction of the target constellation mission intent feature library formed by clustering training data, which includes the target constellation mission intents defined in step S12. As new data is accumulated and confirmed, potential mission intents can be added to the intent library to realize the incremental extensional learning of the model and expand the target constellation mission intent feature set.

[0047] The specific steps of step S4 are as follows: S41. Load the pre-trained convolutional neural network of the feature extractor and the target constellation mission intent feature library into the onboard computer of our spacecraft. S42. For the acquired target star cluster observation data, convert it into a three-dimensional lattice sequence of the target star cluster configuration according to the method described in step S13. S43. Input the three-dimensional lattice sequence of the target star cluster configuration into the feature extractor to extract the target star cluster configuration features; S44. Compare the feature vector with the constructed target constellation mission intent feature library, calculate its distance from the center of each intent cluster, and output the target constellation mission intent label and confidence level.

[0048] In this example, the intent recognition algorithm carried in the onboard computer of our spacecraft consists of a pre-trained feature extractor and a target constellation mission intent feature library.

[0049] First, the perception system of our spacecraft outputs the relative positions of each satellite in the target constellation in the LVLH system at a cycle of 1 minute, and estimates the role type of the satellite through payload signal analysis. The system maintains a sliding time window of 31 frames in real time according to the method in step S12, and generates a three-dimensional dot matrix sequence of the target constellation configuration with a shape of 31×256×256×256×1.

[0050] Then, the feature extractor on the onboard computer performs forward computation on the real-time generated three-dimensional point matrix sequence, quickly outputting a 1024-dimensional feature vector. The computational load of this process is much lower than that of model training, meeting the real-time requirements of the satellite.

[0051] The onboard computer maintains a target constellation mission intent feature library in its memory, which is formed by clustering historical data. This library contains feature ranges of known intent clusters such as "tracking", "reconnaissance", and "standby". Real-time feature vectors are quickly matched with the intent library by calculating the distance to the center of each cluster.

[0052] In this embodiment, the system discovers that multiple consecutive feature vectors are densely located within a known cluster of "tracking" intentions. The system integrates the analysis results of nearly 30 minutes and outputs the conclusion "the current intention of the target constellation is: tracking" with high confidence. This information can be transmitted to the ground command and control center in real time, providing key intelligence support for the safety early warning and strategy formulation of our spacecraft.

[0053] In summary, this invention, based on unsupervised learning, enables intent recognition of target star clusters. During neural network training, data augmentation and contrastive learning enable the feature extraction neural network to learn the general features of the three-dimensional lattice sequence of the target star cluster configuration. This is suitable for training large amounts of unlabeled data. Simultaneously, an adaptive density clustering algorithm is used to autonomously mine the potential behavioral structure of the star cluster, making it suitable for handling the problem of emerging group intents. This is more in line with engineering needs. Furthermore, since the neural network training is completed on the ground, only the neural network needs to be loaded onto the satellite, reducing the computing power burden on the onboard computer and accelerating the intent recognition speed.

Claims

1. A target star cluster intent recognition method based on unsupervised learning, characterized in that, Includes the following steps: S1. Construct the target constellation mission intent set; S2. Based on the TS2Vec self-supervised learning framework, a convolutional neural network is built and trained to learn the three-dimensional lattice sequence features of the target star cluster configuration. S3. Based on the DBSCAN density clustering algorithm, group the three-dimensional lattice sequence of target star cluster configuration; S4. Apply the trained algorithm to target constellation intent recognition.

2. The target constellation intent recognition method based on unsupervised learning according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Based on the payload carried by the satellite, simplify the satellite types in the target constellation to main satellite, communication satellite, observation satellite, and functional satellite; S12. Based on the functional division and motion characteristics of satellites, the complex satellite constellation mission intent is abstracted into ten mission intents: encirclement, reconnaissance, tracking, interference, assembly, dispersion, evasion, escape, protection, and standby. S13. Based on the definition of the target constellation mission intent, construct a three-dimensional lattice sequence example of the target constellation mission intent configuration.

3. The target constellation intent recognition method based on unsupervised learning according to claim 2, characterized in that, The satellite type in the target constellation in step S11 is defined as follows: The primary satellite holds a central leadership position in the target constellation, responsible for collecting and analyzing environmental awareness information from all other satellites in the constellation, making decisions and allocating tasks for space missions, and directing the maneuvers, grouping, and space operations of communication satellites, observation satellites, and functional satellites. Communication satellites are responsible for enabling information exchange between the main satellite and communication satellites, between the main satellite and functional satellites, and between observation satellites and functional satellites. When the distance between the main satellite and the observation satellite or functional satellite exceeds the maximum communication distance, or when the number of communications reaches the maximum limit, communication satellites are needed to enable information exchange among satellites in the constellation. The observation satellite is responsible for acquiring the status information of space targets and other environmental information, and uploading the acquired observation information to the main satellite for decision-making. Functional satellites are satellites that actually perform space missions. They carry various payloads required for the mission, such as docking devices, capture devices, reconnaissance devices, and electromagnetic interference devices. They are commanded by the host satellite and perform maneuvers, formations, and space operations. The target constellation mission intent in step S12 is defined as follows: Encirclement refers to a state in which a certain number of target constellation satellites approach our spacecraft from multiple directions and multiple orbital planes, forming a geometric distribution and motion pattern in the space region around our spacecraft. Reconnaissance refers to the state in which a certain number of target satellite constellations pass through the space region surrounding our spacecraft in a relatively flying orbit. Tracking refers to the state in which a certain number of target satellite constellations maintain a certain distance from our spacecraft by taking relative motion orbits, either around or accompanying it. Interference refers to the state in which a certain number of target satellite constellations repeatedly pass through the space region near the transfer orbit of our spacecraft in a short period of time, adopting a relatively fast-moving, skimming trajectory. Gathering refers to the state in which a group of satellites approach a certain location in space by taking a relative motion orbit, either around or accompanying it, and the relative distance between them is no greater than a certain value within a certain period of time. Dispersion refers to a situation where the relative distances between satellites in a target constellation change from a state where the distances remain constant for a long period of time to a state where they generally increase. Evasion refers to the state of actively avoiding a certain space region during the transfer of a target star cluster to a certain location; Escape refers to the state in which the distance between the target constellation satellites and our spacecraft gradually increases; "Escort" refers to the state in which most of the satellites in a target constellation move in relative orbits or fly alongside one or more of the satellites in the target constellation. Standby refers to the state in which the satellites of a target constellation maintain their current relative orbit for an extended period of time; The target constellation configuration three-dimensional lattice sequence in step S13 refers to mapping the position of each satellite in the target constellation in the LVLH coordinate system onto the target constellation configuration three-dimensional lattice map, and then taking three-dimensional lattice maps at equal time intervals to form the target constellation configuration three-dimensional lattice sequence, the data shape of which is as follows: N × x × y × z × s The specific meaning of each dimension of data is as follows: N This refers to the first [number] point in the three-dimensional lattice sequence of the target star cluster configuration. N The frame contains a 3D dot matrix of the target star cluster configuration, with values ​​ranging from 0 to 30. The time interval between adjacent 3D dot matrix images of the target star cluster configuration is 1 minute. x This refers to the satellite's position in the LVLH coordinate system. x The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. x The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. x =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. y This refers to the satellite's position in the LVLH coordinate system. y The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. y The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. y =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. z This refers to the satellite's position in the LVLH coordinate system. z The axial component is mapped to the nearest point on the 3D lattice diagram of the target star cluster configuration. z The axis coordinates are an arithmetic sequence with values ​​ranging from -128 to 127, each interval being 1. z =0 coincides with the origin of the LVLH coordinate system, and the interval between adjacent coordinates in the LVLH coordinate system is 50 meters. s This refers to the type of satellite located at this coordinate on the three-dimensional dot matrix diagram of the target constellation configuration. Different values ​​represent: 0 indicates no satellite at this coordinate, 1 indicates the target constellation's main satellite at this coordinate, 2 indicates the target constellation's communication satellite at this coordinate, 3 indicates the target constellation's observation satellite at this coordinate, 4 indicates the target constellation's functional satellite at this coordinate, and 5 indicates our spacecraft at this coordinate.

4. The target constellation intent recognition method based on unsupervised learning according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Augment the current batch of target star cluster configuration 3D lattice sequence by rotation, translation and cropping to generate positive and negative sample pairs; S22. Construct a convolutional neural network feature extractor. Input the augmented 3D point matrix sequence into the convolutional neural network feature extractor to extract the feature vector of the 3D point matrix sequence. S23. Construct a multilayer perceptron neural network projector, input the feature vector into the multilayer perceptron neural network projector, and output the projection vector. S24. Design a contrastive loss function and train a convolutional neural network based on the projection vector to learn the general features of the three-dimensional lattice sequence of the target star cluster configuration.

5. The target constellation intent recognition method based on unsupervised learning according to claim 4, characterized in that, The data augmentation and positive / negative sample pair generation in step S21 refers to applying a set of random but physically meaningful transformations to generate a "view" of a batch of 3D point matrix sequence samples to form positive sample pairs. These transformations aim to simulate the uncertainty in observation while maintaining the intended semantics. The specific methods of data augmentation are as follows: Spatial rotation refers to a small, random rotation around a certain axis in the LVLH coordinate system to simulate a minute change in the observation perspective; Random translation refers to a small-scale random translation within the entire lattice space; Sequence pruning refers to the random selection of continuous segments in the time dimension; The convolutional neural network feature extractor in step S22 is based on a convolutional neural network design, with the input layer being... N × x × y × z × s The target star cluster configuration three-dimensional point array sequence data is processed through five convolutional layers for adaptive pooling, and then through three fully connected layers to finally output the feature vector of the three-dimensional point array sequence. The multilayer perceptron neural network projector in step S23 is composed of two layers of multilayer perceptron neural network. It maps the feature vectors output by the feature extractor to the contrast learning space, so that the projection vectors of similar images are closer together on the unit hypersphere, providing an optimization target for loss calculation. The expression for the contrast loss function in step S24 is as follows: in, z i and z j These represent two feature vectors of the same positive sample pair. z k The feature vector representing the negative sample. This indicates that the cosine similarity is taken. Indicates temperature hyperparameter, 2 N This indicates the number of data items in a batch.

6. The target constellation intent recognition method based on unsupervised learning according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31. Based on experience and the elbow rule, set the initial parameters of the DBSCAN density clustering algorithm; S32. Input the feature vectors extracted from step S2 into the DBSCAN density clustering algorithm. Based on the local density of the input dataset, discover intention clusters of arbitrary shape and number, and adaptively adjust the algorithm parameters. S33. The clustering results aggregate the features of the three-dimensional lattice sequence of the target star cluster configuration into multiple intent clusters, and assign task intent semantic labels to the intent clusters. S34. Construct and maintain the target constellation mission intent feature set formed by clustering results.

7. The target constellation intent recognition method based on unsupervised learning according to claim 6, characterized in that, In step S31, the initial parameters of the DBSCAN density clustering algorithm are the neighborhood radius. ε The minimum number of neighborhood samples, MinPts, for the core points; The specific steps of step S32 are as follows: S321. Based on experience, set a smaller [value]. ε initial value ε 0 and a fixed MinPts value; S322, will ε Inputting 0 as a candidate value into DBSCAN performs clustering to obtain the number of clusters. C 0; S323. Increase the candidate value by a fixed step size. σ The number of clusters is obtained by clustering again. C 1; S324, Repeat step S323; S325. Monitor changes in the number of clusters; if a preset number is detected... n Continuous candidate values ε k , ε k+1 , ..., ε n Number of clusters generated C k , C k+1 ,,…, C n If all candidate values ​​are identical, the clustering result is considered to have entered a relatively stable plateau region, and the largest value among these candidate values ​​is selected as the final result. ε parameter; In step S33, assigning task intent semantic labels to intent clusters means that by associating them with the target constellation task intents defined in step S12, each stable cluster with a large amount of data can be assigned a corresponding semantic label. For feature vectors that cannot be classified into any existing cluster or a new, small and dense cluster, the system can identify them as potential unknown intents, trigger an alarm, and hand them over to experts for analysis. The construction and maintenance of the target constellation mission intent feature set formed by the clustering results in step S34 refers to the construction of the target constellation mission intent feature library formed by clustering training data, which includes the target constellation mission intents defined in step S12. As new data is accumulated and confirmed, potential mission intents can be added to the intent library to realize the incremental extensional learning of the model and expand the target constellation mission intent feature set.

8. The target constellation intent recognition method based on unsupervised learning according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Load the pre-trained convolutional neural network of the feature extractor and the target constellation mission intent feature library into the onboard computer of our spacecraft. S42. For the acquired target star cluster observation data, convert it into a three-dimensional lattice sequence of the target star cluster configuration according to the method described in step S13. S43. Input the three-dimensional lattice sequence of the target star cluster configuration into the feature extractor to extract the target star cluster configuration features; S44. Compare the feature vector with the constructed target constellation mission intent feature library, calculate its distance from the center of each intent cluster, and output the target constellation mission intent label and confidence level.

Images