Unmanned cluster identification method based on motion topology reasoning

Abstract

The invention specifically relates to an unmanned cluster identification method based on motion topology reasoning, and belongs to the technical field of artificial intelligence and unmanned systems. The method comprises the following steps: firstly, collecting a historical track sequence of unmanned cluster nodes, and constructing an interpretable feature system containing three types of physical features of motion, interaction and topology; inputting the feature vector into a self-adaptive encoder containing dynamic graph construction, multi-head attention and time sequence aggregation components, and outputting a high-order cluster feature vector of time sequence aggregation; then, target identity recognition and node interaction topology reasoning are completed in sequence, and a cluster node future trajectory prediction result is output through a dual decoder; and finally, identifying and predicting performance through joint optimization of a multi-objective loss function, and updating model parameters in real time to adapt to a dynamic task scene. According to the method, the problems of poor scale adaptability, task decoupling and feature black box of an existing method are effectively solved, and end-to-end joint optimization of identity recognition and trajectory prediction is realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and unmanned systems technology, specifically relating to an unmanned swarm identification method based on motion topology reasoning, which is applicable to variable-scale unmanned swarm targets in complex dynamic environments and can accurately identify target labels. Background Technology

[0002] With the development of unmanned systems technology, large-scale unmanned swarms are increasingly widely used in various fields. Among them, target identification is the core technology to ensure the safety of swarm collaboration and avoid accidental attacks and conflicts.

[0003] Existing unmanned swarm identification methods mainly fall into two categories: one relies on traditional signal features, with the core being identification based on the target's features or signals from a dedicated identifier. This type of method has significant limitations in future large-scale unmanned swarm scenarios: first, it is highly dependent on dedicated equipment, leading to a surge in hardware costs during large-scale swarm deployments, and small unmanned platforms struggle to accommodate complex sensing modules; second, it has weak anti-interference capabilities, with signal features easily distorted or tampered with, resulting in a sharp drop in identification accuracy; and third, it lacks rigidity in target differentiation, as global manufacturing and platform models converge, making effective identification difficult solely based on signal features. The other category, motion pattern recognition methods based on machine learning, while freeing them from reliance on signal features, is still underdeveloped: currently, there are few relevant methods, and models use fixed-structure networks to extract trajectory features, failing to fully explore the physical laws within motion patterns, resulting in a lack of interpretability for the features and a "black box"-like model decision-making process; furthermore, these methods often process identification and swarm trajectory independently, ignoring the coupling relationship between the two—identity labels determine prior motion patterns, while motion patterns provide the basis for identity determination, leading to isolated task execution and limited overall performance.

[0004] From theoretical research to engineering implementation, existing unmanned swarm recognition technologies still face many common problems that urgently need to be addressed, particularly in complex environments: First, they lack adaptability to cooperative targets. Most methods are based on modeling conventional motion trajectories and do not fully consider the cooperative maneuvers performed by swarm targets to achieve tasks such as detection, leading to significantly increased model bias in recognizing and predicting unexpected motion patterns. Second, model accuracy is limited in dynamic environments. The performance of existing solutions highly depends on stable perception conditions. In complex scenarios, the stability of feature extraction decreases, making it difficult to guarantee model recognition and prediction accuracy. Third, they are weak in handling dynamic changes in swarm nodes. Swarm splitting, merging, and other node number changes are common in swarm tasks, but existing network architectures are mostly designed with fixed dimensions, lacking sufficient generalization ability to changes in node scale, requiring retraining to adapt to new scenarios. Fourth, there is an imbalance between interpretability and practicality. Deep learning-based solutions often rely on automatic feature extraction by the model, resulting in "black box" decision-making with ambiguous physical meaning of features. Furthermore, the design of core parameters lacks clear basis, making model optimization difficult and hindering engineering verification. These problems severely restrict the reliable application of unmanned swarm recognition technology in real-world scenarios.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides an unmanned cluster recognition method based on motion topology reasoning. By constructing a physically interpretable cluster feature system, designing an adaptive encoder architecture, and establishing a collaborative optimization mechanism for topology reasoning and recognition, it solves the core problems in the prior art such as poor scale adaptability, task decoupling, and feature black box, and achieves end-to-end joint optimization of trajectory prediction and recognition.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to a first aspect of the present invention, a method for identifying unmanned clusters based on motion topology reasoning is provided, the method comprising: Step 1: Collect historical trajectory data of unmanned swarms and extract three types of physical features, including motion features, interaction features, and topological features; based on swarm dynamics rules, encode the interaction relationships between individuals into graph structure edge types, and construct a motion topology graph by combining the individual's own state and environmental constraints. The motion topology graph serves as the input for subsequent feature encoding. Step 2: Construct an adaptive encoder that includes node encoding units, edge encoding units, temporal fusion units, and attention units. The adaptive encoder performs feature mapping, temporal fusion, and joint encoding on the physical features and motion topology graph from Step 1, and outputs a high-order cluster feature vector of temporal aggregation. Step 3: Based on the high-order cluster feature vector of temporal aggregation, construct an end-to-end model of topological reasoning, feature extraction and target recognition by combining the variational reasoning framework, design a multi-objective joint loss function, update the model parameters through backpropagation, and realize the collaborative optimization of topological reasoning, identity recognition and trajectory prediction. Step 4: Input the historical trajectory of the unmanned cluster to be identified into the trained end-to-end model, and output the target attribute determination result and topology evolution matrix.

[0009] In some exemplary embodiments, the motion features described in step 1 include velocity consistency, velocity-acceleration features, and trajectory smoothness; the interaction features include interaction stability, relative distance, and relative velocity; and the topological features include formation dispersion and shape entropy.

[0010] In some exemplary embodiments, step 2, which involves performing feature mapping, temporal fusion, and joint encoding on the physical features and motion topology graph from step 1 using the adaptive encoder to output a high-order cluster feature vector of temporal aggregation, specifically involves: The node encoding unit and the edge encoding unit use a graph neural network (GNN) to perform feature mapping on the node state and edge type of the motion topology graph, and introduce an adaptive normalization layer to dynamically adjust the normalization parameters according to the number of input nodes. The temporal fusion unit employs a bidirectional long short-term memory network (LSTM) to capture the dynamic evolution characteristics of the topology graph from both forward and reverse temporal directions. The attention unit combines a multi-head attention mechanism to jointly encode physical features and dynamic evolution features of the topology graph, and outputs a high-order cluster feature vector of temporal aggregation.

[0011] In some exemplary embodiments, the node encoding unit and the edge encoding unit adopt a 3-layer graph neural network with a hidden layer dimension of 128 and an activation function of exponential linear unit (ELU). The edge encoding unit calculates the edge embedding through dynamic node-to-edge conversion and sets four parallel attention heads to learn the weight distribution of different interaction rules.

[0012] In some exemplary embodiments, the temporal fusion unit employs a 2-layer bidirectional LSTM with a hidden layer dimension of 64, capturing topological evolution features of 100 time steps from both forward and reverse temporal directions.

[0013] In some exemplary embodiments, step 3, which involves constructing an end-to-end model for topological reasoning, feature extraction, and target recognition based on the high-order cluster feature vector of temporal aggregation, combined with a variational reasoning framework, specifically includes: The prior and posterior distributions of cluster features are calculated by using the temporally aggregated high-order cluster feature vectors output by the encoder. Discrete topology inference results are obtained by using Gumbel-Softmax sampling, thereby realizing real-time dynamic inference of motion topology. The topological inference results are fused with the cluster global features and then input into a fully connected layer and a Softmax classifier to output the target attribute probability. The cluster global features are obtained by feature dimensionality reduction and global pooling processing from the temporal aggregated high-order cluster feature vector output by the adaptive encoder. A dual-decoder architecture is adopted, which outputs the future trajectory prediction results of each cluster node based on the recognition results and topological inference features respectively.

[0014] In some exemplary embodiments, the multi-objective joint loss function mentioned in step 3 is a joint loss function of topology reconstruction loss, divergence loss, cluster feature loss, and target recognition loss; The topology reconstruction loss is calculated by the cross-entropy between the inference topology and the true topology; the divergence loss uses Kullback-Leibler divergence; the cluster feature loss uses mean squared error (MSE) loss; and the target identification loss uses cross-entropy loss. The weight ratio of each loss is adjusted by balancing the weight coefficients to prioritize classification accuracy and feature consistency.

[0015] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the unmanned cluster identification method based on motion topology reasoning described in the first aspect above.

[0016] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the unmanned cluster identification method based on motion topology reasoning described in the first aspect above.

[0017] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the motion topology reasoning-based unmanned cluster identification method described in the first aspect above by executing the executable instructions.

[0018] The unmanned swarm identification method based on motion topology reasoning provided by the embodiments of the present invention addresses the core defects of traditional methods in the background art by constructing a physically interpretable feature system, designing an adaptive encoder architecture, and a collaborative optimization mechanism guided by topology reasoning. It has outstanding technical advantages and practical value: Compared with traditional methods, the present invention only extracts physical features based on the historical trajectory data of the swarm to achieve target identification, completely eliminating the dependence on dedicated equipment, significantly reducing the hardware cost of large-scale swarm deployment, and adapting to various small unmanned platforms.

[0019] To address the weakness of traditional signal feature-based methods in resisting interference, this invention uses the physical laws of cluster motion and topological interactions as the core of identification, mitigating the impact of complex environments and maintaining stable identification accuracy even in complex scenarios. It also overcomes the bottleneck of insufficient target differentiation rigidity caused by similar equipment models, achieving accurate identification by exploring essential differences such as motion coordination and formation structure. Compared to existing machine learning-based motion pattern recognition methods, this invention constructs a feature system with clear physical meaning, including speed consistency and formation dispersion, solving the "black box" problem. This makes the model's recognition logic interpretable and provides a basis for core parameter optimization, significantly improving engineering verification efficiency. By establishing an end-to-end collaborative optimization mechanism, it leverages the coupling relationship between identity tags and motion patterns to improve overall performance. Furthermore, the dual-decoder architecture can simultaneously output target identity and future trajectory prediction results, providing comprehensive support for cluster collaborative avoidance and task planning. Furthermore, the adaptive encoder designed in this invention can automatically adjust model parameters according to the number of input nodes, and can cope with dynamic changes in node scale such as formation splitting and fusion without retraining. At the same time, it can capture the cluster collaborative maneuvering mode and reduce the recognition and prediction deviation in unexpected motion scenarios. The system has a simple architecture, standardized interfaces, strong compatibility, and low deployment cost, and has extremely high engineering practicality and practical value.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 This is a schematic diagram of the overall framework of the unmanned cluster identification method of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the structure of the interpretable feature system of the present invention.

[0024] Figure 3 This is a schematic diagram of the internal architecture of the adaptive encoder of the present invention.

[0025] Figure 4 This is a flowchart of the trajectory prediction process using the dual decoder of the present invention.

[0026] Figure 5 This is a schematic diagram of the topology-guided collaborative optimization mechanism of the present invention.

[0027] Figure 6 This is a schematic diagram of the unmanned cluster identification method of the present invention. Detailed Implementation

[0028] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0029] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0030] To address the shortcomings and deficiencies of existing technologies, this example implementation provides an unmanned cluster recognition method based on motion topology reasoning. By constructing a physically interpretable feature system, designing an adaptive encoder architecture that supports variable-scale clusters, and implementing a multi-task collaborative optimization mechanism guided by topology reasoning, it achieves high-precision, robust, and highly interpretable unmanned cluster recognition in complex dynamic environments.

[0031] A method for identifying unmanned swarms based on motion topology reasoning is provided. The method includes: an interpretable swarm physical feature system, an adaptive encoder, a topology reasoning-guided collaborative optimization mechanism, and an identification process; the above key links are progressively integrated to form a complete technical chain.

[0032] The interpretable cluster physical feature system is responsible for collecting historical trajectory data of unmanned clusters and extracting interpretable features with clear physical meaning; The adaptive encoder is responsible for adapting to the feature extraction requirements of variable-sized clusters and completing the feature and recognition label encoding. The topology-guided collaborative optimization mechanism is responsible for combining the decoder to achieve end-to-end joint optimization of multiple tasks, including cluster trajectory prediction and cluster identification. The identification process is responsible for inputting the historical trajectory data of the unmanned cluster to be identified into the trained model, and outputting the dynamic target identification label.

[0033] Among them, the interpretable features include three categories and eight types of interpretable features with clear physical meaning, including motion features, interaction features, and topological features. Motion features include velocity consistency, velocity-acceleration features, and trajectory smoothness; interaction features include interaction stability, relative distance, and relative velocity; and topological features include formation dispersion and shape entropy. The innovation of the feature system of this invention lies in the fact that existing technologies only use a small number of dimensional features in isolation, which cannot fully depict the essential laws of cluster cooperative motion. Motion features reflect the synergy between individuals and groups, interaction features characterize the dynamic correlation strength between nodes, and topological features quantify the overall structural characteristics of the cluster. These three types of features are progressive and mutually corroborative, forming a full-dimensional representation of individuals, node interactions, and the group. Compared to existing abstract features without physical meaning, the three types of features in this feature system are all designed based on group dynamics rules, which not only ensures the interpretability of the features but also accurately captures the essential differences between different clusters, providing a clear physical basis for subsequent identification. On this basis, the interaction relationships between individuals are encoded into graph structure edge types based on group dynamics rules. Combining the above physical features with the individual's own state and environmental constraints, a dynamically updated motion topology graph is constructed to realize the structured representation of the cluster motion pattern.

[0034] The adaptive encoder is constructed by including node encoding units, edge encoding units, temporal fusion units, and attention units. The node encoding units and edge encoding units use GNN graph neural networks to perform feature mapping on node states and edge types. The edge encoding unit learns the weight distribution of different interaction rules through parallel attention heads. To eliminate the feature distribution offset caused by scale differences, an adaptive normalization layer is introduced, which can dynamically adjust the normalization parameters according to the number of input nodes. The temporal fusion unit uses a bidirectional long short-term memory network to capture the dynamic evolution features of the topology graph from both forward and reverse time directions, retaining motion mode information at multiple time steps. The attention unit combines a multi-head attention mechanism to jointly encode local and global features of the cluster, strengthen the feature representation of key motion modes, and suppress redundant information interference.

[0035] Subsequently, based on the variational inference framework, an end-to-end model of "topology inference, feature extraction, and target recognition" is constructed. The prior and posterior distributions of cluster features are calculated by outputting features from the encoder. Discrete topology inference results are obtained using Gumbel-Softmax sampling, enabling real-time dynamic inference of moving topology. The topology inference results are fused with the global features of the cluster and input into a fully connected layer and a Softmax classifier, outputting the target attribute probability of each node. To address the coupling relationship between identity recognition and trajectory prediction, a dual-decoder architecture is designed, corresponding to each cluster, and outputting future trajectory prediction results based on the recognition results and topology features respectively. At the same time, a multi-objective joint loss function is constructed, including topology reconstruction loss, divergence loss, cluster feature loss, and target recognition loss. The weight ratio of each loss is adjusted by a balancing coefficient to prioritize classification performance and feature consistency. Finally, the overall model parameters are updated through backpropagation.

[0036] Finally, the target identification and result output are completed. The historical trajectory data of the unmanned swarm to be identified is input into the trained model. Dynamic features are extracted through the adaptive encoder. After processing by the topology reasoning and collaborative optimization module, three core results are output simultaneously: the target attribute determination results of each node, the topology evolution matrix (representing the dynamic changes in the interaction relationship of the swarm), and the future trajectory prediction results of the nodes. The model supports real-time parameter updates and can flexibly adapt to changes in motion patterns caused by complex environments and changes in task types.

[0037] This invention also protects an unmanned swarm identification system applying the above-described method. This system includes a data acquisition module, a topology modeling module, an adaptive coding module, a collaborative reasoning module, and a result output module. Each module's function is adapted to the core aspects of the above-described method. Specifically, the data acquisition module collects historical trajectory data of the unmanned swarm and extracts physical state features such as individual position and velocity, completing data fusion and preprocessing, and supporting multi-source sensor data fusion. The topology modeling module constructs a motion topology graph and interpretable swarm physical features based on swarm dynamics rules, outputting a feature matrix and a topology graph adjacency matrix. The adaptive coding module performs the feature extraction process, adaptively adapting to the dynamic feature representation of swarms of different sizes through dynamic node-to-edge and edge-to-node transformations and normalization layers, achieving node / edge feature mapping, temporal fusion, and key feature enhancement. The collaborative reasoning module performs topology reasoning, identification, and trajectory prediction, optimizing model performance through a joint loss function, and simultaneously completing motion topology dynamic reasoning, target identification, and future node trajectory prediction. The result output module outputs the target attribute determination result, the topology evolution matrix, and the future trajectory prediction result.

[0038] refer to Figure 1 and Figure 6As shown, this invention unfolds in four main stages: feature system construction, adaptive encoding, collaborative optimization, and result output. Specifically, it may include the following steps: Step 1: Collect historical trajectory data of unmanned swarms and extract three types of physical features, including motion features, interaction features, and topological features; based on swarm dynamics rules, encode the interaction relationships between individuals into graph structure edge types, and construct a motion topology graph by combining the individual's own state and environmental constraints. The motion topology graph serves as the input for subsequent feature encoding. Step 2: Construct an adaptive encoder that includes node encoding units, edge encoding units, temporal fusion units, and attention units. The adaptive encoder performs feature mapping, temporal fusion, and joint encoding on the physical features and motion topology graph from Step 1, and outputs a high-order cluster feature vector of temporal aggregation. Step 3: Based on the high-order cluster feature vector of temporal aggregation, construct an end-to-end model of topological reasoning, feature extraction and target recognition by combining the variational reasoning framework, design a multi-objective joint loss function, update the model parameters through backpropagation, and realize the collaborative optimization of topological reasoning, identity recognition and trajectory prediction. Step 4: Input the historical trajectory of the unmanned cluster to be identified into the trained end-to-end model, and output the target attribute determination result and topology evolution matrix.

[0039] The following will describe each step in more detail with reference to the accompanying drawings and specific embodiments.

[0040] In step 1, an interpretable cluster physical feature system is constructed. Historical trajectory data of unmanned clusters are collected, and three types of physical features are extracted: cluster motion features, interaction features, and topological features. Based on the group dynamics rules, the interaction relationships between individuals are encoded as graph structure edge types. Considering the individual's own state and environmental constraints, a motion topology graph is constructed. This graph serves as the core input of the node encoding unit and edge encoding unit in the adaptive encoder, providing the interaction structure foundation between cluster nodes for the graph neural network. This topology graph is used for feature mapping of node states and edge types, and at the same time, it provides structured data support for the temporal fusion unit to capture the dynamic evolution features of the topology graph.

[0041] Specifically, the interpretable physical features mentioned in step 1 are defined based on group features: motion features include velocity consistency, velocity-acceleration features, and trajectory smoothness; interaction features include interaction stability, relative distance, and relative velocity; topological features include formation dispersion and shape entropy. All features have clear physical meanings, forming a multi-dimensional interpretable feature representation of individuals, interactions, and groups.

[0042] For example, based on the collected historical trajectory data of unmanned swarms, eight types of interpretable features need to be extracted, combined with... Figure 2The feature system shown is structured as follows, and the specific calculation methods for each feature are as follows: (1) Speed ​​Consistency: The Euclidean distance between the individual speed and the group average speed is calculated and normalized to reflect the degree of coordination between the individual and the group's movement, and to quantify the degree of deviation between the individual movement and the group average movement. The calculation formula is as follows:

[0043] in, For a moment No. The velocity vector of each node For a moment The average velocity vector of all nodes in the cluster. For Euclidean distance operator, This is the normalization factor for the maximum value.

[0044] (2) Velocity-acceleration characteristics: The acceleration modulus is calculated by the velocity difference between adjacent time steps, reflecting the rate of change of the nodal motion state and characterizing the rapid or gradual change of motion. The calculation formula is as follows:

[0045] in, For a moment No. The velocity vector of each node For time step.

[0046] (3) Trajectory smoothness: It is measured by the continuity of the directions of adjacent displacement vectors. The smaller the angle, the smoother the trajectory. The calculation formula is:

[0047] in, For a moment No. The position vectors of each node. To avoid tiny values ​​that are divided by zero.

[0048] (4) Interaction stability: Characterized by the inverse relationship of the magnitude of the rate of change of velocity. The smaller the rate of change of velocity, the more stable the interaction. This measures the fluctuation of the interaction intensity between an individual and its neighbors. The calculation formula is as follows:

[0049] in, For a moment No. The velocity vector of each node.

[0050] (5) Relative distance: The average distance between an individual and its nearest neighbor is taken and the maximum value is limited. This reflects the spatial proximity between nodes. The calculation formula is as follows:

[0051] in, This is the upper limit threshold for distance (set according to the cluster task scenario). For the first The neighboring nodes of a node.

[0052] (6) Relative velocity: The magnitude of the velocity difference between an individual and its nearest neighbor is calculated, reflecting the coordination and intensity of movement between nodes. The calculation formula is as follows:

[0053] in, For the first The nearest neighbor index of each node.

[0054] (7) Formation dispersion: This refers to the distance distribution of individuals from the group center, reflecting the density of the cluster and the deviation of individuals from the group centroid. The calculation formula is as follows:

[0055] in, For a moment The center location vector of the cluster This represents the number of nodes in the cluster.

[0056] (8) Shape entropy: Quantified by the entropy value of the histogram of relative position distribution. The smaller the entropy value, the more regular the shape of the cluster. It quantifies the irregularity of the local formation structure observed from the individual perspective. The calculation formula is:

[0057] in, The number of bins in the histogram. For the first The relative distance of a node to other nodes in the th order The probability in each bin Used to avoid zero in logarithmic operations.

[0058] In addition, the construction of the motion topology graph needs to be based on the group dynamics rules. Interactions between individuals whose relative distance is less than a set threshold and whose relative speed meets the cooperation conditions are encoded as strong interaction edges. Other interactions that meet the basic communication range are encoded as weak interaction edges. No interaction relationship is encoded as an edge type. The topology graph is dynamically updated by combining the above features with individual position, speed status and environmental constraints (such as obstacle range).

[0059] In step 2, an adaptive encoder architecture for a variable-scale cluster is designed, constructing an adaptive encoder comprising node encoding units, edge encoding units, temporal fusion units, and attention units. The node and edge encoding units use a graph neural network (GNN) to perform feature mapping on the node states and edge types encoded based on group dynamics rules in the motion topology graph constructed in step 1, and introduce an adaptive normalization layer to dynamically adjust the normalization parameters according to the number of input nodes. The temporal fusion uses a bidirectional long short-term memory (LSTM) network to capture the dynamic evolution features of the topology graph. The attention unit combines a multi-head attention mechanism to jointly encode the physical features and dynamic evolution features of the topology graph in step 1, outputting a high-order cluster feature vector of temporal aggregation. This feature vector serves as the core input feature for topological reasoning and target recognition in the next step, providing a feature basis for calculating the prior / posterior distribution of cluster features and realizing recognition decisions.

[0060] Specifically, in step 2, the edge encoding unit achieves variable-scale adaptation by dynamically calculating the edge embedding between nodes, and the parallel attention head learns the weight distribution of different interaction rules; the temporal fusion unit captures the dynamic evolution features of the topology graph from the forward and reverse time directions, and outputs the temporal aggregation features fused from multiple time points; the attention unit strengthens the representation of key features through a multi-head self-attention mechanism, and realizes the joint encoding of local and global features of the cluster.

[0061] For example, refer to Figure 3 The architecture design employs a 3-layer graph neural network for both the node encoding unit and the edge encoding unit, with a hidden layer dimension of 128 and an Exponential Linear Unit (ELU) activation function. The edge encoding unit calculates edge embeddings through dynamic node-to-edge transformation and uses four parallel attention heads to learn the weight distribution of different interaction rules. The adaptive normalization layer dynamically adjusts the statistical parameters of batch normalization based on the number of input nodes to eliminate feature distribution shifts caused by variations in node size. The temporal fusion unit uses a 2-layer bidirectional LSTM with a hidden layer dimension of 64, capturing topological evolution features over 100 time steps in both forward and reverse temporal directions. The attention unit uses four attention heads to perform weighted fusion of local and global features, enhancing the feature response of key motion patterns.

[0062] In step 3, a collaborative optimization mechanism for cluster trajectory and recognition guided by topological reasoning is established. Based on the variational reasoning framework, an end-to-end model of "topological reasoning, feature extraction, and target recognition" is constructed. A joint loss function including topological reconstruction loss, divergence loss, cluster feature loss, and target recognition loss is designed, and the model parameters are updated through backpropagation. Specifically, step 1 is the basic layer, which completes the extraction of cluster physical features and the construction of motion topology graph, providing physical features and structured topological input for subsequent feature encoding; step 2 is the feature encoding layer, which performs adaptive feature extraction and fusion on the input of step 1, and outputs high-order temporal aggregated features; step 3 is the optimization reasoning layer, which realizes end-to-end modeling of topological reasoning, target recognition, and trajectory prediction based on the output features of step 2, and optimizes the feature system and encoding model parameters constructed in steps 1 and 2 through the joint loss function. The three are in a progressive logical relationship of data input, feature encoding, and reasoning optimization.

[0063] Specifically, in step 3, the prior and posterior distributions of cluster features are calculated using the temporal aggregated features output by the encoder. Discrete topology inference results are obtained using Gumbel-Softmax sampling, enabling real-time dynamic inference of the motion topology. The recognition and decision module fuses the topology inference results with the global cluster features, inputs them into a fully connected layer and a Softmax classifier, and outputs the target attribute probability. A dual-decoder architecture is adopted, outputting the future trajectory prediction results for each cluster node based on the recognition results and the topology inference features. The global cluster features are obtained by feature dimensionality reduction and global pooling processing from the temporal aggregated high-order cluster feature vector output by the adaptive encoder. The true values ​​of the cluster features used to calculate the loss during model training are calculated from the position, velocity, and trajectory information of the real unmanned cluster nodes according to the formulas in the specific implementation details below.

[0064] Specifically, in step 3, the topology reconstruction loss is calculated by the cross-entropy between the inference topology and the real topology. The divergence loss uses Kullback-Leibler (KL) divergence, the cluster feature loss uses mean squared error (MSE) loss, and the target identification loss uses cross-entropy loss. The weight ratio of each loss is adjusted by balancing the weight coefficients to prioritize classification accuracy and feature consistency.

[0065] For example, in combination Figure 4The schematic diagram shown illustrates a principle based on a variational inference framework. Both the prior and posterior distributions are modeled using Gaussian distributions. The temperature parameter sampled by Gumbel-Softmax is initially set to 1.0 during training and linearly reduced to 0.1 with each training iteration (200 rounds), balancing sampling randomness and stability. The classifier employs two fully connected layers with a hidden layer dimension of 64, using a Rectified Linear Unit (ReLU) activation function, and a Softmax activation function for the output layer. Both decoders utilize a message passing network + 3-layer MLP structure, as shown in the reference diagram. Figure 5 The trajectory prediction process uses a decoder hidden layer dimension of 64 and a dropout probability of 0.1. In the joint loss function, the topology reconstruction loss uses cross-entropy loss, the divergence loss uses KL divergence, the cluster feature loss uses MSE loss, and the recognition loss uses cross-entropy loss. The balance coefficients of each loss are set to 1.0 for topology reconstruction loss, 0.5 for divergence loss, 5.0 for cluster feature loss, and 10.0 for recognition loss, respectively. This weight allocation prioritizes classification accuracy.

[0066] In step 4, target identification and result output are achieved. The historical trajectory of the unmanned cluster to be identified is input into the trained model, and the target attribute determination result and topology evolution matrix are output.

[0067] Specifically, combined Figure 1 The overall process involves post-processing the model output, determining the target attribute with a probability threshold of 0.5, outputting the topology evolution matrix in the form of an adjacency matrix, and predicting the future trajectory to output the position and velocity information for subsequent time steps. It supports real-time updates of model parameters, with the update frequency dynamically adjusted according to the environment.

[0068] Example Based on the technical solution described in this invention, specific application scenarios are given below, and the practicality and effectiveness of this invention are further illustrated with the accompanying drawings. This embodiment focuses on the task of "dynamic identification and trajectory prediction of a 10-node hybrid cluster," highlighting the implementation process and core advantages of the method.

[0069] 1. Data Acquisition and Preprocessing This embodiment collects motion trajectory data of an unmanned swarm in a collaborative scenario using sensors. This includes 100 sets of data from our own swarm and 100 sets of data from the target swarm. Data dimensions include timestamps (sampling frequency 10Hz), individual IDs, position coordinates (unit: meters), and velocity vectors (unit: meters / second). The data preprocessing process is as follows: the data is denoised, missing values ​​are filled using linear interpolation, and the data is divided into training, validation, and test sets in a 6:2:2 ratio. The training set is used for model training, the validation set is used to validate and update the model, and the test set is used for performance testing. After preprocessing, the data is then processed according to… Figure 2 The feature system shown calculates various interpretable features and constructs a motion topology graph as model input data. This data consists of two parts: a "feature matrix (dimension: time step × number of nodes × number of cluster features)" and a "topology graph adjacency matrix (dimension: time step × number of nodes × number of nodes)," which fully reflects the physical characteristics and interaction structure of cluster motion.

[0070] 2. Model Training Configuration The model is implemented using the PyTorch deep learning framework, and the core parameter configurations strictly follow the relevant code design to ensure project reproducibility. (Reference) Figure 3 Adaptive encoder architecture and Figure 4 The collaborative optimization mechanism has the following core parameter configurations: (1) Adaptive encoder: The node encoding unit and the edge encoding unit adopt a 3-layer GNN with a hidden layer dimension of 128 and an activation function ELU; the dynamic node-to-edge and edge-to-node transformation adapts to the variable node scale, and 4 parallel attention heads learn the interaction rule weights; the adaptive normalization layer adjusts the batch normalization parameters according to the number of nodes; the temporal fusion unit is a 2-layer bidirectional LSTM with a hidden layer dimension of 64, capturing the topological evolution features of 100 time steps; the attention unit has 4 attention heads, which weightedly fuse local and global features.

[0071] (2) Collaborative reasoning module: Based on the variational reasoning framework, the prior and posterior distributions are Gaussian distributions; the Gumbel-Softmax sampling temperature parameter is linearly reduced from 1.0 to 0.1; the recognition classifier is a 2-layer fully connected layer (64 hidden layers, ReLU activation) + Softmax output; the dual decoder adopts a message passing network and a 3-layer MLP, with a hidden layer dimension of 64 and dropout=0.1.

[0072] (3) Training parameters: batch size = 32, number of iterations = 200, the optimizer is Adaptive Moment Estimation (Adam), the initial learning rate = 1e-4, which decays to 0.5 every 50 rounds; the joint loss function weights are configured as 1.0 (topology reconstruction): 0.5 (KL divergence): 5.0 (feature loss): 10.0 (recognition loss) to prioritize classification accuracy; an early stopping strategy is adopted (training is stopped if the accuracy of the validation set does not improve for 10 consecutive rounds) to prevent overfitting.

[0073] During training, the model dynamically extracts cluster features of all nodes through an adaptive encoder, identifies the differences in the interaction structure between the two parties through a topology inference module, and learns the motion patterns of the two types of clusters through dual decoders, thereby achieving end-to-end optimization of "feature extraction, topology inference, target recognition, and trajectory prediction".

[0074] 3. Test Scenario Setup and Result Verification The test scenario simulates a complex environment: the hybrid cluster is initially in a queuing and assembly state (initial 5 seconds), and 10% signal noise and a 5% packet loss rate are added to the scenario to simulate a complex environment. Implementation process combined... Figure 1 The overall framework is as follows: (1) Input data: Input the preprocessed trajectory data obtained into the trained model, including the position and velocity of all nodes; (2) Adaptive encoding: The model processes the interaction relationship of nodes through dynamic transformation, the adaptive normalization layer eliminates the feature offset caused by the node size, the bidirectional LSTM captures the topological evolution features of the formation and aggregation, and the attention unit strengthens the key features. (3) Collaborative reasoning: Based on the variational reasoning framework, Gumbel-Softmax sampling infers the dynamic topology; the target recognition module combines topological features and physical features to output the label probability of each node; the dual decoders predict the trajectories of the cluster nodes of both sides according to the recognition results. (4) Output results: Output three types of core results, including target attribute labels, topology evolution matrix, and future trajectory data.

[0075] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0076] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0077] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A method for identifying unmanned clusters based on motion topology reasoning, characterized in that, The method includes: Step 1: Collect historical trajectory data of unmanned swarms and extract three types of physical features, including motion features, interaction features, and topological features; based on swarm dynamics rules, encode the interaction relationships between individuals into graph structure edge types, and construct a motion topology graph by combining the individual's own state and environmental constraints. The motion topology graph serves as the input for subsequent feature encoding. Step 2: Construct an adaptive encoder that includes node encoding units, edge encoding units, temporal fusion units, and attention units. The adaptive encoder performs feature mapping, temporal fusion, and joint encoding on the physical features and motion topology graph from Step 1, and outputs a high-order cluster feature vector of temporal aggregation. Step 3: Based on the high-order cluster feature vector of temporal aggregation, construct an end-to-end model of topological reasoning, feature extraction and target recognition by combining the variational reasoning framework, design a multi-objective joint loss function, update the model parameters through backpropagation, and realize the collaborative optimization of topological reasoning, identity recognition and trajectory prediction. Step 4: Input the historical trajectory of the unmanned cluster to be identified into the trained end-to-end model, and output the target attribute determination result and topology evolution matrix.

2. The unmanned cluster identification method based on motion topology reasoning according to claim 1, characterized in that, The motion features mentioned in step 1 include velocity consistency, velocity-acceleration features, and trajectory smoothness; the interaction features include interaction stability, relative distance, and relative velocity; and the topological features include formation dispersion and shape entropy.

3. The unmanned cluster identification method based on motion topology reasoning according to claim 1, characterized in that, Step 2, which describes using the adaptive encoder to perform feature mapping, temporal fusion, and joint encoding on the physical features and motion topology graph from Step 1, outputting a high-order cluster feature vector of temporal aggregation, specifically involves: The node encoding unit and the edge encoding unit use a graph neural network (GNN) to perform feature mapping on the node state and edge type of the motion topology graph, and introduce an adaptive normalization layer to dynamically adjust the normalization parameters according to the number of input nodes. The temporal fusion unit employs a bidirectional long short-term memory network (LSTM) to capture the dynamic evolution characteristics of the topology graph from both forward and reverse temporal directions. The attention unit combines a multi-head attention mechanism to jointly encode physical features and dynamic evolution features of the topology graph, and outputs a high-order cluster feature vector of temporal aggregation.

4. The unmanned cluster identification method based on motion topology reasoning according to claim 3, characterized in that, The node encoding unit and the edge encoding unit adopt a 3-layer graph neural network with a hidden layer dimension of 128 and an activation function of exponential linear unit (ELU). The edge encoding unit calculates the edge embedding through dynamic node-to-edge transformation and sets four parallel attention heads to learn the weight distribution of different interaction rules.

5. The unmanned cluster identification method based on motion topology reasoning according to claim 3, characterized in that, The temporal fusion unit uses a 2-layer bidirectional LSTM with a hidden layer dimension of 64 to capture the topological evolution features of 100 time steps in both forward and reverse temporal directions.

6. The unmanned cluster identification method based on motion topology reasoning according to claim 1, characterized in that, Step 3, which involves constructing an end-to-end model for topological reasoning, feature extraction, and target recognition based on high-order cluster feature vectors derived from temporal aggregation and combined with a variational reasoning framework, specifically includes: The prior and posterior distributions of cluster features are calculated by using the temporally aggregated high-order cluster feature vectors output by the encoder. Discrete topology inference results are obtained by using Gumbel-Softmax sampling, thereby realizing real-time dynamic inference of motion topology. The topological inference results are fused with the cluster global features and then input into a fully connected layer and a Softmax classifier to output the target attribute probability. The cluster global features are obtained by feature dimensionality reduction and global pooling processing from the temporal aggregated high-order cluster feature vector output by the adaptive encoder. A dual-decoder architecture is adopted, which outputs the future trajectory prediction results of each cluster node based on the recognition results and topological inference features.

7. The unmanned cluster identification method based on motion topology reasoning according to claim 1, characterized in that, The multi-objective joint loss function mentioned in step 3 is a joint loss function of topology reconstruction loss, divergence loss, cluster feature loss, and target recognition loss; The topology reconstruction loss is calculated by the cross-entropy between the inference topology and the true topology; the divergence loss uses Kullback-Leibler divergence; the cluster feature loss uses mean squared error (MSE) loss; and the target identification loss uses cross-entropy loss. The weight ratio of each loss is adjusted by balancing the weight coefficients to prioritize classification accuracy and feature consistency.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the unmanned cluster identification method based on motion topology reasoning as described in any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the unmanned cluster identification method based on motion topology reasoning as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the unmanned cluster identification method based on motion topology reasoning according to any one of claims 1 to 7 by executing the executable instructions.

Images