Methods Of Combat Formation Identification Based On Graph Neural Networks

Under the background of information war,the saying "know yourself and know your enemy,and you can fight a hundred battles with no danger of defeat" is still important.The explosive growth of big data makes the identification of enemy combat formations by traditional methods seriously affect our decision-making efficiency and reduce the possibility of victory.Faced with massive incomplete,untimely,inaccurate and even wrong battlefield information with deceptive,commanders find it difficult to analyze the hidden key information.Therefore,this paper conducts a series of researches on battle formation recognition based on graph neural network,and proposes a set of methods and analysis procedures for battle formation recognition,which mainly solves three problems: fragmented information processing,node role recognition and intention recognition.Aiming at the problem of information processing of combat formation fragmentation,in order to process the fragmented information,make it suitable for solving by graph neural network and remove some redundant information,this paper proposes an algorithm of generating battle formation map based on Motif structure.This method includes Motif discovery,Motif extraction,Motif matching and Motif merging algorithms.In the process of extraction and recombination,redundant information and nodes in the actual battle formation map are removed,so as to excavate the important information in the fragmented information and master the possible intention types of the battle formation,so as to support the subsequent identification of the battle formation.In this paper,the role recognition of combat formation nodes is abstracted as node classification based on graph neural network,and multi-kernel graph neural network is used to classify nodes and infer node attributes of combat formation,which can describe the network of weapon equipment system more accurately.By observing the network structure and node attributes to predict the role types of entities,the task types and combat intentions of combat formations can be more accurately judged.Moreover,in order to make the classification results more reliable and more consistent with the real situation,the network reliability of the role identification of the nodes in the combat formation based on Graph neural network is studied,and the Graph Calibration Loss(GCL)is proposed to calibrate the reliability of the model.The reliability of prediction is consistent with the real situation,and the decision makers can trust the results of model prediction more.Specifically,the reliability of output confidence of the model is improved by model calibration of graph neural network.It encourages the network to assign high weights to samples with high confidence.The loss is regularized by cross entropy loss and predicted distribution entropy.Optimizing the cross entropy can ensure the high precision of the model,while optimizing the predicted distribution entropy,i.e.minimizing the entropy,can increase the confidence of the model output.A large number of experimental results show that GCL model calibration effect is excellent,and can ensure high precision in the training process.In this paper,the problem of battle formation intention recognition is abstracted as a graph classification problem based on graph neural network,and adaptive graph pooling neural network is used to capture the global representation of battle formation diagram,and the battle formation intention recognition is carried out.Using adaptive graph pooling neural network to calculate the probability of various possible operational intentions of the composite graph,the most likely operational intentions can be obtained.Finally,we use the simulated aircraft carrier formation data for case analysis,which proves the effectiveness of the method and analysis process for combat formation identification.