Search

CN-122021931-A - Unmanned cluster recognition method based on motion topology reasoning

CN122021931ACN 122021931 ACN122021931 ACN 122021931ACN-122021931-A

Abstract

The invention particularly relates to an unmanned cluster recognition method based on motion topology reasoning, and belongs to the technical field of artificial intelligence and unmanned systems. The method comprises the steps of firstly collecting historical track sequences of unmanned cluster nodes, constructing an interpretable characteristic system containing 3 types of physical characteristics of motion, interaction and topology, inputting the interpretable characteristic system into a self-adaptive encoder containing dynamic graph construction, multi-head attention and time sequence aggregation components, outputting high-order cluster characteristic vectors of time sequence aggregation, then completing target identity recognition and node interaction topology reasoning, outputting future track prediction results of the cluster nodes through double decoders, finally optimizing recognition and prediction performance through multi-target loss function combination, and updating model parameters in real time to adapt to dynamic task scenes. The method effectively solves the problems of poor scale adaptability, task decoupling and characteristic black box of the existing method, and realizes end-to-end joint optimization of identity recognition and track prediction.

Inventors

  • YANG SHUHENG
  • ZHANG DONG
  • DENG JIE
  • LIU WENYI

Assignees

  • 西北工业大学

Dates

Publication Date
20260512
Application Date
20260312

Claims (10)

  1. 1. An unmanned cluster recognition method based on motion topology reasoning, which is characterized by comprising the following steps: Step 1, collecting historical track data of an unmanned cluster, extracting 3 types of physical characteristics comprising motion characteristics, interaction characteristics and topological characteristics, encoding inter-individual interaction relations into graph structure edge types based on group dynamics rules, and constructing a motion topological graph by combining individual states and environmental constraints, wherein the motion topological graph is used as input of subsequent characteristic codes; Step 2, constructing an adaptive encoder comprising a node coding unit, an edge coding unit, a time sequence fusion unit and an attention unit, and performing feature mapping, time sequence fusion and joint coding on the physical features and the motion topological graph in the step 1 through the adaptive encoder to output a time sequence aggregated high-order cluster feature vector; step 3, constructing an end-to-end model of topology reasoning, feature extraction and target identification by combining a variation reasoning frame based on a high-order cluster feature vector of time sequence aggregation, designing a multi-target joint loss function, and updating model parameters through back propagation to realize collaborative optimization of the topology reasoning, the identity identification and the track prediction; and 4, inputting the historical track of the unmanned cluster to be identified into a trained end-to-end model, and outputting a target party attribute judging result and a topology evolution matrix.
  2. 2. The unmanned cluster recognition method based on motion topology reasoning, which is disclosed in claim 1, is characterized in that the motion features in step 1 comprise speed consistency, speed-acceleration features and track smoothness, the interaction features comprise interaction stability, relative distance and relative speed, and the topology features comprise formation dispersibility and shape entropy.
  3. 3. The unmanned cluster recognition method based on motion topology reasoning according to claim 1, wherein in step 2, the adaptive encoder performs feature mapping, time sequence fusion and joint encoding on the physical features and the motion topology of step 1, and outputs a high-order cluster feature vector of time sequence aggregation, which specifically comprises: the node coding unit and the edge coding unit adopt a graph neural network GNN to perform feature mapping on the node state and the edge type of the motion topological graph, an adaptive normalization layer is introduced, and normalization parameters are dynamically adjusted according to the number of input nodes; the time sequence fusion unit adopts a two-way long-short-term memory network LSTM to capture dynamic evolution characteristics of the topological graph from the time positive sequence and the reverse sequence directions; the attention unit combines a multi-head attention mechanism to perform joint coding on physical characteristics and topological graph dynamic evolution characteristics, and outputs high-order cluster characteristic vectors of time sequence aggregation.
  4. 4. The unmanned cluster recognition method based on motion topology reasoning according to claim 3, wherein the node coding unit and the edge coding unit adopt a 3-layer graph neural network, the hidden layer dimension is set to 128, the activation function adopts an exponential linear unit ELU, the edge coding unit calculates and embeds the edge through dynamic node-to-edge conversion, and 4 parallel attention heads are set to learn weight distribution of different interaction rules.
  5. 5. The unmanned cluster recognition method based on motion topology reasoning of claim 3, wherein the timing fusion unit adopts 2-layer bidirectional LSTM, the hidden layer dimension is set to 64, and the topology evolution characteristics of 100 time steps are captured from the time positive sequence and the reverse sequence directions.
  6. 6. The unmanned cluster recognition method based on motion topology reasoning according to claim 1, wherein the high-order cluster feature vector based on time sequence and time sequence aggregation in the step 3 is combined with a variation reasoning frame to construct an end-to-end model of topology reasoning, feature extraction and target recognition, and the method is specifically as follows: Calculating prior distribution and posterior distribution of cluster features through time sequence aggregated high-order cluster feature vectors output by an encoder, and obtaining a discrete topology reasoning result by utilizing Gumbel-Softmax sampling to realize real-time dynamic reasoning of motion topology; The topological reasoning result and the cluster global feature are fused and then input into a full-connection layer and Softmax classifier to output the attribute probability of the target party, wherein the cluster global feature is obtained by carrying out feature dimension reduction and global pooling on the time sequence aggregation Gao Jieji cluster feature vectors output by the self-adaptive encoder; and adopting a double-decoder architecture, and respectively outputting future track prediction results of all cluster nodes based on the identification results and the topological reasoning features.
  7. 7. The unmanned cluster recognition method based on motion topology reasoning according to claim 1, wherein the multi-objective joint loss function in step 3 is a joint loss function of topology reconstruction loss, divergence loss, cluster feature loss and objective recognition loss; The method comprises the steps of calculating the cross entropy of a reasoning topology and a real topology according to the topology reconstruction loss, adopting a Kullback-Leibler divergence for the divergence loss, adopting a mean square error MSE loss for the cluster characteristic loss, adopting a cross entropy loss for the target party identification loss, adjusting the weight duty ratio of each loss through a balance weight coefficient, and preferentially guaranteeing the classification precision and the characteristic consistency.
  8. 8. A storage medium having stored thereon a computer program, which when executed by a processor implements the unmanned cluster recognition method based on motion topology reasoning as claimed in any of claims 1 to 7.
  9. 9. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the unmanned cluster recognition method based on motion topology reasoning as claimed in any one of claims 1 to 7.
  10. 10. An electronic device, comprising: Processor, and A memory for storing executable instructions of the processor; Wherein the processor is configured to perform the unmanned cluster recognition method based on motion topology reasoning of any of claims 1 to 7 via execution of the executable instructions.

Description

Unmanned cluster recognition method based on motion topology reasoning Technical Field The invention belongs to the technical field of artificial intelligence and unmanned systems, and particularly relates to an unmanned cluster recognition method based on motion topology reasoning, which is suitable for variable-scale unmanned cluster targets in a complex dynamic environment and can accurately recognize target party labels. Background Along with the development of unmanned system technology, the large-scale unmanned clusters are increasingly widely applied in various fields, wherein target party identification is a core technology for guaranteeing the cooperative security of the clusters and avoiding false hits and conflicts. The existing unmanned cluster recognition method is mainly divided into two types, one type relies on traditional signal characteristics, and the core is to perform identity judgment by acquiring characteristics of a target or a special recognizer signal. The method has the large limitations in future large-scale unmanned cluster scenes that firstly, the special equipment dependence is high, the hardware cost is increased rapidly when the large-scale unmanned cluster is deployed, a small unmanned platform is difficult to bear a complex sensing module, secondly, the anti-interference capability is weak, signal characteristics are easy to distort or falsify, so that the recognition accuracy is suddenly reduced, thirdly, the rigidity of target party distinction is insufficient, the model of the platform is converged along with global popularization of manufacturing, and the effective discrimination is difficult only through the signal characteristics. The other type of motion pattern recognition method based on machine learning gets rid of dependence on signal characteristics, but has imperfect development, namely the current related method has few track characteristics, a model adopts a fixed structure network to extract track characteristics, the physical rule in the motion pattern is not fully mined, the characteristic lacks interpretability, the model decision process is like a black box, the method is often used for independently processing recognition and cluster tracks, the coupling relation between the recognition and cluster tracks, namely an identity tag decides the prior of the motion pattern, the motion pattern provides basis for identity judgment, and the task is isolated and executed, so that the overall performance is limited. From theoretical research to engineering landing, the existing unmanned cluster recognition technology still has a plurality of common problems to be solved, the problems are particularly prominent in complex environments, firstly, the adaptation capability of cooperative targets is insufficient, most methods are based on conventional motion track modeling, cooperative maneuver implemented by the cluster targets for realizing tasks such as detection is not fully considered, so that the recognition and prediction deviation of the model on unexpected motion modes is obviously increased, secondly, the model precision is limited in the dynamic environment, the performance of the existing scheme highly depends on stable perception conditions, the stability of feature extraction is reduced in the complex scenes, the model recognition and prediction precision is difficult to guarantee, thirdly, the dynamic change of cluster nodes is weak, the increase and decrease of the number of nodes such as formation splitting and fusion are normal to the cluster tasks, but the existing network architecture is mostly of fixed dimension design, the generalization capability of the node scale change is insufficient, new scenes can be adapted only by retraining, fourthly, the method interpretability and practicability are unbalanced, the scheme based on deep learning is used for automatically extracting features, so that ' black box ' decisions ' are formed, the physical meaning of features is fuzzy, and the core parameter design is difficult to realize, and the verification of the engineering is difficult to be realized. These problems severely restrict the reliable application of unmanned cluster recognition technology in practical scenarios. It should be noted that the information disclosed in the above background section is only for enhancing understanding of the background of the invention and thus may include information that does not form the prior art that is already known to those of ordinary skill in the art. Disclosure of Invention The invention provides an unmanned cluster recognition method based on motion topology reasoning, which solves the core problems of poor scale suitability, task decoupling, feature black box and the like in the prior art by constructing a physical interpretable cluster feature system, designing a self-adaptive encoder architecture and establishing a cooperative optimization mechanism of topology reasoning and recognition, and realizes